Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Design of a Wearable Two-Dimensional Joystick as
a Muscle-Machine Interface Using
Mechanomyographic Signals
by
Deba Pratim Saha
Thesis submitted to the faculty of the
Virginia Polytechnic Institute and State University
in partial fulfillment of the requirements for the degree of
Master of Science
In
Computer Engineering
Thomas L. Martin
Ico I. Bukvic
R. Benjamin Knapp
November 12, 2013
Blacksburg, VA
Keywords: Gesture Recognition, Wearable Joystick, Mechanomyography, Pattern Recognition.
Design of a Wearable Two-Dimensional Joystick as a Muscle-Machine Interface
Using Mechanomyographic Signals
by
Deba Pratim Saha
ABSTRACT
Finger gesture recognition using glove-like interfaces are very accurate for sensing individual
finger positions by employing a gamut of sensors. However, for the same reason, they are also
very costly, cumbersome and unaesthetic for use in artistic scenarios such as gesture based music
composition platforms like Virginia Tech’s Linux Laptop Orchestra. Wearable computing has
shown promising results in increasing portability as well as enhancing proprioceptive perception
of the wearers’ body. In this thesis, we present the proof-of-concept for designing a novel muscle-
machine interface for interpreting human thumb motion as a 2-dimensional joystick employing
mechanomyographic signals. Infrared camera based systems such as Microsoft Digits and
ultrasound sensor based systems such as Chirp Microsystems’ Chirp gesture recognizers are
elegant solutions, but have line-of-sight sensing limitations. Here, we present a low-cost and
wearable joystick designed as a wristband which captures muscle sounds, also called
mechanomyographic signals. The interface learns from user’s thumb gestures and finally interprets
these motions as one of the four kinds of thumb movements. We obtained an overall classification
accuracy of 81.5% for all motions and 90.5% on a modified metric. Results obtained from the user
study indicate that mechanomyography based wearable thumb-joystick is a feasible design idea
worthy of further study.
iii
Acknowledgements
I want to take this opportunity to extend my heartfelt thanks to my thesis advisor Dr. Tom Martin
for his continued encouragement and guidance throughout my work. His feedbacks have helped
shape me as an individual. His ideas have helped shape this work in its current form. I would also
like to thank him for diligently reviewing my thesis write up from scratch and bearing with my
not-so-good compositions.
I want to express my sincere gratitude to Dr. Ben Knapp for supporting me throughout the thesis
work with invaluable ideas and feedbacks. His words of wisdom in numerous interactions we had
throughout the duration of this thesis work has provided me with invaluable insights into the world
of research and professionalism.
I am grateful to Dr. Ico Bukvic for supporting me throughout my Masters’ program. I want to
thank him for providing with the opportunity to work on this challenging problem statement, for
being extraordinarily understanding with my minor slip-ups and getting me involved with Institute
for Creativity, Arts and Technology at Virginia Tech. I am also indebted to him for his help during
my minor accident.
I also thank DISIS and ICAT at Virginia Tech for supporting me financially throughout the
duration of my Master’s program. All the photographs included in this thesis are captured by the
author.
I would like to extend my heartfelt thanks to my friend Prithwish Chakraborty for being a constant
mentor and always being with me as a brother.
I would also like to express my gratitude to my wonderful friends - Ishan Mehta, Saurav Ghosh,
Arijit Chattopadhyay, Rakesh Sehgal, Avinash Desai, Manpreet Hora, Alex Foldenaeur and
Candice Foldenaeur in Blacksburg and Anant Goenka, Kushal Lakhotia, Dolonchapa Chakraborty,
Saurav Gupta, Debopam D. Chaudhury, my cousin brother Rahul Saha, Rangeen B.R. Choudhury,
Ayush Kedia, and many more whose presence helped me sail through this entire period. Thanks
for being around me and lending your patient ears whenever I needed you all.
Thanking your own parents is never enough, still it shouldn’t stop one. I would like to thank my
Ma, Mrs. Maya Saha and Baba, Mr. Akhil Ranjan Saha for all the sacrifices they have made.
Thanks to my elder brother Mr. Partha Pratim Saha for being a constant source of inspiration.
If not for the constant encouragement from my best friend, philosopher and guide, Ms. Sanchita
Roy, it wouldn’t have been possible to conclude this work. Thanking her will be an understatement,
but I’ll always be indebted to her for being there and brightening me up whenever needed.
Deba Pratim Saha
Blacksburg, Virginia, USA.
November 12, 2013.
iv
Table of Contents
Chapter 1 – Introduction ........................................................................................................................... 1
1.1 Motivation ......................................................................................................................................... 1
1.2 Problem Statement ........................................................................................................................... 2
1.3 Methodology ..................................................................................................................................... 3
1.4 Contributions .................................................................................................................................... 6
1.5 Thesis Organization ........................................................................................................................... 6
Chapter 2 – Background and Prior Work ................................................................................................... 7
2.1 Finger Gesture Recognition .............................................................................................................. 7
2.1.1 Optical Linear Encoder based Sensing ...................................................................................... 8
2.1.2 Ultra Sound based Sensing........................................................................................................ 8
2.1.3 Microsoft Digits – Wrist Worn IR Based Finger Sensing ........................................................... 9
2.2 Muscle Machine Interfaces (MMI) .................................................................................................. 10
2.2.1 Surface Electromyography and Its Limitations ....................................................................... 10
2.2.2 Mechanomyography as a Better MMI Signal .......................................................................... 11
2.3 Science of Surface Mechanomyography (MMG) ............................................................................ 12
2.3.1 Bio-Mechanics of MMG .......................................................................................................... 12
2.3.2 Characteristics of MMG .......................................................................................................... 13
2.3.3 MMG Control Signal in MMI ................................................................................................... 15
2.4 Thumb Physiology and Sensor Placement ...................................................................................... 16
2.4.1 Thumb Motion as a Joystick Controller .................................................................................. 16
2.4.2 Thumb Muscles and Sensor Placement .................................................................................. 17
2.5 Conclusions ..................................................................................................................................... 21
Chapter 3 - System Overview and Sensor Design .................................................................................... 22
3.1 System Overview............................................................................................................................. 22
3.2 mmgSensor Design ......................................................................................................................... 26
3.2.1 Transducer Selection ............................................................................................................... 26
3.2.2 Design Details of ECM based mmgSensor .............................................................................. 28
3.2.3 Chamber Design ...................................................................................................................... 29
3.3 Sensor Interfacing and Data Acquisition ......................................................................................... 32
3.4 Conclusion ....................................................................................................................................... 33
Chapter 4 - Data Analysis and Pattern Recognition ................................................................................. 35
4.1 Auto-Event Annotation (AEA) Algorithm ........................................................................................ 35
4.2 Sample Feature Extraction .............................................................................................................. 42
4.2.1 Feature-Set Description .......................................................................................................... 42
4.2.2 Curse of Dimensionality .......................................................................................................... 45
4.3 Dimension Reduction and Feature Selection .................................................................................. 47
4.3.1 Pre-selection using Fisher Discriminant Score Analysis .......................................................... 47
v
4.3.2 Principal Components Analysis (PCA) ..................................................................................... 47
4.4 Pattern Recognition ........................................................................................................................ 49
4.4.1 Quadratic Discriminant Analysis (QDA) .................................................................................. 50
4.4.2 k-Nearest Neighbor (kNN) ....................................................................................................... 51
4.4.3 Cross Validation of Test Samples ............................................................................................ 51
4.5 Conclusion ....................................................................................................................................... 52
Chapter 5 - User Study and Results ......................................................................................................... 53
5.1 Experiment and Methods ............................................................................................................... 53
5.1.1 Experimental Protocol Design ................................................................................................. 53
5.2 Results ............................................................................................................................................. 56
5.2.1 Metric Used for Comparing Results ........................................................................................ 56
5.2.2 Choosing Dimensionality for this MMG Dataset..................................................................... 56
5.2.3 Patterns Classification of Thumb Movement Using QDA ....................................................... 58
5.2.4 Patterns Classification of Thumb Movement Using k-NN....................................................... 61
5.3 Discussion ........................................................................................................................................ 65
5.3.1 Discussion on Final Results ..................................................................................................... 65
5.3.2 Discussion on User C ............................................................................................................... 67
Chapter 6 - Conclusion and Future Work ................................................................................................ 68
6.1 Conclusion ....................................................................................................................................... 68
6.2 Future Work .................................................................................................................................... 69
Bibliography ................................................................................................................................................ 70
vi
List of Figures
Figure 1.1 Flow Diagram of the System ........................................................................................................ 5
Figure 2.1 Snapshot of Sensor Placement over Flexor Pollicis Longus muscle. ........................................ 18
Figure 2.2 Snapshot of Sensor Placement over Extensor Pollicis Longus muscle. ..................................... 20
Figure 3.1 Flow Diagram of Major Subsystems ......................................................................................... 23
Figure 3.2 System Block Diagram of Major Subsystems ........................................................................... 24
Figure 3.3 Side view, Top view and Angular view of the sensor respectively. .......................................... 30
Figure 3.4 Snapshot of the Wristband with two embedded mmgSensor. .................................................... 31
Figure 3.5 Line Diagram of the Sensor Housing ......................................................................................... 32
Figure 3.6 Flow Diagram of the Data Acquisition Process ......................................................................... 33
Figure 4.1 MMG capture for four different kinds of thumb motions .......................................................... 36
Figure 4.2 MMG Signal (Blue) and its Energy (Green) Plots for Channel 1 .............................................. 38
Figure 4.3 Aggregate Window Energy (Red) Plot overlaid for Channel 1 ................................................. 38
Figure 4.4 Local Energy Peak (Red) Plots for Channel 1 ........................................................................... 39
Figure 4.5 Global Energy Peaks (Blue dots) for Channel 1 ........................................................................ 39
Figure 4.6 Event Selection among channels (Blue Dots) ............................................................................ 40
Figure 4.7 Final Event Demarcation for each Movement. .......................................................................... 41
Figure 5.1 Snapshots of Neutral, Abduction and Adduction positions (L-to-R) ......................................... 55
Figure 5.2 Snapshots of Flexion and Extension positions (L-to-R) ............................................................ 55
Figure 5.3 True Rate of Classification for the entire Dataset for 2<D<5 (D ∈ 𝐼) ....................................... 57
Figure 5.4 Classification Accuracy per-User per-Class employing QDA Algorithm ................................. 61
Figure 5.5 Classification Accuracies per-User obtained by k-Nearest Neighbor Algorithm ...................... 62
Figure 5.6 True Rate of Classification for the entire Dataset for 1<k<10 (k ∈ 𝐼). ...................................... 63
Figure 5.7 Classification Accuracy per-User per-Class employing k-NN Algorithm ................................. 65
vii
List of Tables
Table 2.1 Tabular representation of Thumb Movements and Its Muscles ................................................. 19
Table 3.1 Functional Subsystems in which the current work is divided. .................................................... 23
Table 4.1 Table listing the parameters assumed to be prior knowledge...................................................... 36
Table 5.1 Participant-wise Classification Accuracies using PCA and QDA Algorithm ............................. 58
Table 5.2 Participant-wise Classification Accuracies using Fisher Score, PCA and QDA Algorithm. ...... 60
Table 5.3 Participant-wise Classification Accuracies using PCA and k-NN Algorithm ............................ 64
Table 5.4 Summary of Results for 2 Classifiers using the Two Metrics Described in Section 5.2.1 .......... 66
1
Chapter 1 – Introduction Traditional musical instruments are some of the most fascinating, elegant and sophisticated control
interfaces developed by human beings [1]. The complexity and depth of control provided by acoustic
musical instruments such as a piano or a cello is impressive. The impetus towards the development of novel
musical instruments has been a potent driver of innovative human machine interfaces. Due to the need for
musical instruments to have complex control, yet a simple man-machine interface as well as reliable design,
they are often at the forefront of development of man-machine control interfaces [1].
Examples of various kinds of alternative musical interfaces using different forms of control signals are
strewn across in research literature. Bio-signal based and gesture based musical interfaces are two such
examples of alternative interfaces for musical instruments. Earliest use of bio-signals in music generation
can be traced back to Alvin Lucier’s Music for Solo Performer (1965) [2]. This interface senses the brain
wave signals generated during a particular mental state of the performer, to control percussion instruments
[3]. Bio-sensing technologies have been studied in great details in recent years, with a notable biophysical-
only music performances being developed by a team of researchers at SARC (Sonic Arts and Research
Center) in Belfast, led by the main researcher of Bio Muse Project, Dr. Benjamin Knapp [4].
1.1 Motivation
Similar to bio-signal based control, multimodal gesture based interfaces for music composition have also
been explored in recent years. A notable few such musical ensembles are Princeton Laptop Orchestra
(PLOrk), Stanford Laptop Orchestra (SLOrk) and Virginia Tech’s one-of-a-kind Linux Laptop Orchestra
(L2-Ork). The unique property of these setups is that they use human gestural interactions with some form
of computing infrastructure to alter the tone, pitch, tempo and other parameters of music during a
performance. The computing infrastructure as mentioned above can be an ensemble of laptops being
manipulated by a group of artists as in the case of SLOrk [5] and Wii-game controller remotes, also called
2
Wii-motes and Nunchucks, being controlled by a group of performers in case of L2-Ork [6]. Linux Laptop
Orchestra or L2-Ork, developed at Virginia Tech by Dr. Ico Bukvic and Dr. Tom Martin, is a unique step
towards a complete gesture based music composition suite. It is also unique in the sense that, due to its
gesture based music composition abilities, it has an additional aspect of music choreography in each music
piece. Many gestures that are used in L2Ork’s performances for composing music are adapted from the art
form called Tai-Chi (or Taiji) [7].
In a typical L2Ork performance, a group of performers hold a Wii-mote with one hand and Nunchuck with
the other. The hand gesture data are picked up from the Wii-mote and Nunchuck connected to a netbook
via Bluetooth. The input data is processed for gestural inputs which is used to compose music. As described
above, L2-Ork incorporates Tai-Chi movements in its choreography which requires extremely fluidic body
movements and gestures while maintaining a relaxed state of body and mind [8, 9]. For the performer’s
body to be in a relaxed state and the hand movements to be freeform fluidic, we postulate that while
performing the Tai-Chi moves performer’s hands should be in an open and relaxed position. However, this
is not possible in the current L2-Ork system due to the need for conscious gripping of Wii-motes and
Nunchucks by the performers during the entire duration of performance. Although this current system is at
a disadvantage in this aspect, but, the data acquired from the Wii-mote and Nunchuck is rich in gestural
information. In addition, the sheer abundance of control modalities available on these Wii-controllers makes
them indispensable for the L2-Ork system. Thus, the new interface must be very similar in functionality
and control as compared to Wii-mote and Nunchucks.
1.2 Problem Statement
The problem with the current controller used in L2-Ork as stated above, prompted us towards designing a
grip-free gestural controller which is similar in performance to Wii-controllers in acquiring data and has
comparable number of control inputs to Wii-controllers, if not more. To recognize complex hand gestures,
researchers have applied many possible sensor technologies such as resistive-ink flex sensor in Power
3
Glove, thin foil strain gauges in CyberGlove designed by Stanford University and three space-magnetic
tracker in SpaceGlove designed by Virtuality Entertainment Systems [10]. There are many commercially
available data gloves too such as 5DT, Peregrene and VHand [11]. Although, these gloves effectively solve
the problem of collecting gestural data from the user hand and fingers, they are either prohibitively costly
or too cumbersome and fragile [10]. In addition, they all lack in one common functionality – the
conventional joystick, which is present on Nunchucks and is heavily used in the L2-Ork system, thus are
indispensable for its performance. Voyles et al. have shown the design of a magnetic sensor based wearable
joystick in [10]. However, they are using the user’s wrist movements as direction input to emulate computer
joystick. We cannot use wrist movement to interpret as joystick movements in the L2Ork system due to the
inherent mapping of wrist yaw, roll and pitch inputs from Wii-motes to already existing L2Ork commands.
The Nunchuck joystick has a separate functionality in L2-Ork and cannot be mapped to wrist motions. Thus
our problem statement is to design a low cost and reliable wearable user interface that can interpret some
user movement as a 2 dimensional joystick. As will be shown in details in Chapter 2 that we found the
thumb movement to be most suited for interpreting as joystick movement.
1.3 Methodology
In this thesis, we try to address the above mentioned problem by emulating human thumb movements as a
conventional joystick controller. Our goal is to design a novel wearable thumb joystick that can emulate
the 2-dimensional movements of a conventional joystick present on the Wii-Nunchucks. The biomechanical
feasibility of mapping thumb movements as a 2-dimensional joystick controller has been discussed in
details in Chapter 2. The sensing method described by Voyles et al. uses a rare-earth bar magnet fixed to
the moving part of the hand. The exact location of this magnet is sensed using four giant-magneto-resistant
(GMR) sensors placed underneath the bar magnet on the forearm. This setup is simple and effective in
sensing large movements such as that of the wrists, yet cumbersome for sensing small movements of the
thumb. To narrow down our search for economically viable and reliable sensors that are well suited for
4
wearable applications, we initially developed some prototypes using carbonized stretch sensors arranged in
a grid-like formation. In another such initial prototype, we used four bend sensors on four sides of the thumb
to detect its motion. These prototypes, though simple, didn’t show much promise for reliably sensing thumb
motions and interpreting them as a computer joystick.
Muscle machine interfaces have shown promising results in sensing human motions. A detailed discussion
on muscle-based interfaces has been presented in Chapter 2. As will be elaborated in Chapter 2, we decided
to select one type of bio-signal called Mechanomyogram (MMG) or muscle-sounds for design of this
wearable joystick. Mechanomyogram is a low frequency mechanical vibration occurring at the skin-surface
just above the muscle site which is responsible for making the particular motion [12, 13]. MMG can be
acquired using very low cost microphone modules [14]. Owing to their simple and highly robust design,
they are well suited in the wearable scenarios such as ours. Hence, the final goal of this thesis is to design
a low cost, reliable and wearable muscle machine interface to interpret thumb movement as a joystick using
mechanomyographic signals.
5
Figure 1.1 Flow Diagram of the System
A flow diagram of the system developed for this thesis is shown above in Fig 1.1. The system consists of a
wearable wristband to be strapped on the forearm of the user. The wristband has two mmgSensors embedded
in it which is connected to an interfacing and data acquisition circuitry. A Windows laptop is used for the
pattern-recognition and data analysis of the MMG data. A future development roadmap for this system will
consist of the real-time implementation of this system as well as development of a standalone wearable
application to be used for live performances.
The system developed for implementing this thesis can be broadly divided into four subsystems – MMG
signal acquisition; auto event sampling from the time-series data; feature extraction and feature selection;
and finally pattern recognition on MMG data. As will be shown in Chapter 3 and Chapter 4, the major work
done as a part of this thesis is the development of the mmgSensor, the development of the auto event
sampling algorithm and exploring a set of expressive features of MMG pattern recognition.
6
1.4 Contributions
This work presents a novel approach towards design of a wearable joystick using the thumb movements.
As described earlier in brief in Section 1.2, prior methods developed by Voyle et al. are insufficient as a
solution for our problem statement. We will discuss in further details in Chapter 2, muscle machine
interfaces, and mechanomyography in particular, has some inherent advantages associated with it when
used to design control interfaces. To the best of our knowledge, MMG based pattern recognition has not
been tried on this particular set of thumb motions and we are the first to report good accuracy on thumb
motion recognition. In addition, as will be discussed in Chapter 4, we have developed a novel signal energy
based auto-event annotation algorithm for annotating events in the MMG time-series which will be useful
in designing a fully automated real-time gesture recognition system.
1.5 Thesis Organization
The following chapters will provide an in-depth analysis of the design of mechanomyography based low
cost wearable thumb-joystick. Chapter 2 provides further background on the biomechanics of human thumb
movement and muscles of forearm as well as overviews the technicalities of mechanomyography. Chapter
3 presents a detailed analysis of the sensor design, transducer selection, and chamber design for the
mechanomyographic data acquisition. MMG feature extraction, feature reduction, pattern recognition and
various other data analysis nuances have been dealt with in details in Chapter 4. Details of the user study,
data collection methodology, data analysis and final pattern recognition results are presented in Chapter 5.
Finally, Chapter 6 concludes the work with a short discussion on future work.
7
Chapter 2 – Background and Prior Work This chapter presents an overview of the finger gesture recognition problem statement which we aim to
address in this thesis. It gives a brief summary of currently available technologies used for efficiently
recognizing finger gestures and their shortcomings viz-a-viz our motivation for designing a
mechanomyography based muscle machine interface. This chapter also discusses the justification for our
choice of the control signal used for the system i.e., mechanomyography and its appropriateness in our use
case scenario.
2.1 Finger Gesture Recognition
The fingers of the human hand provide a set of highly precise and intricate mechanisms for interacting with
the environment. Due to the need for control interfaces for human-computer interaction (HCI) to be precise,
accurate and highly reliable, our main mode of interaction for modern machine interfaces are through our
hands and its fingers. For gesture-based interfaces to be effective, it is critical for them to be able to
determine the actual gestures as accurately as possible. This is a difficult task and remains an area of active
research [15].
The problem of precise finger movement detection in three dimensions (3D) can be visualized as a
combination of multiple angle detection problems at each joint. However, current interfaces rarely leverage
the full dexterity of human hands, largely due to the technical challenges encountered in complete detection
of finger motions in 3D, with its many degrees-of-freedom. Consequently, a constrained set of finger
motions along different dimensions have been used for HCI such as tracking 2D inputs only [16, 17],
fingertips and other specific parts [18], supporting interactions through surfaces [19]. Particularly in
computer vision community, researchers are starting to solve the real-time 3D finger pose recognition [20,
21]. Here we present a small discussion on some of the already used technologies used for finger gesture
based human computer interactions.
8
2.1.1 Optical Linear Encoder based Sensing
Technologies such as inertial measurement unit (IMU) based motion tracking systems and mechanical
tracking that are used often in full body motion analysis, could be used for finger pose recognition too.
Even though they are quite accurate, they suffer from the problems of poor wearability and excessive sensor
fusion. They are also expensive, need long and precise calibration sessions and accumulate errors (such as
from gyroscopes) over time [22]. Optical Linear Encoding (OLE) is an optical version of high resolution
position measurement apparatuses called Linear Encoders used to measure position by converting it into
linear motion on fitted scale. Nguyen et al. in [22] have shown the development of an OLE based wrist
movement detection system that is nonintrusive, non-cumbersome and is designed with the long term
wearability considerations in mind. A similar OLE-based finger joint pose detection enclosure for the
fingers would be one possibility for our problem statement, but due to the need for one OLE sensor per
bend for each finger on both hands, the instrumentation tends to become cumbersome and costly. There
would have been huge space constraints for the enclosure for the hand too. Moreover, we did not want to
have a glove based system because of wearability and aesthetic considerations. Thus this method was
deemed unfit for our purpose.
2.1.2 Ultra Sound based Sensing
Commercial computer vision based systems are capable of sensing coarse finger gestures such as the
Toshiba Qosmio G55 laptop uses its front facing camera to control PowerPoint slides [15]. But they have
an inherent problem of sensitivity to lighting conditions and need a lot of processing. Xbox Kinect is another
vision based system which can detect coarse body gestures in 3D. However, Kinect SDK does not support
finger gesture recognition [23]. In addition, vision based systems have the problem of occlusion. As an
alternative, sonic finger gesture recognition using an ultrasound transducer and the Doppler effect for
motion sensing has been shown to work for reliably sensing in-air finger gestures [15, 24, 25]. The speed,
direction and amplitude of the sound waves received were used to detect a wide range of gestures. This
approach does not have the problem of sensitivity to lighting conditions inherent in vision based systems.
9
However, for minute finger pose sensing, occlusion and line-of-sight limitations may present problems in
this type of sensing too. The sensitivity drops rapidly with increasing distance from the sensors and
ultrasound based designs have inherent noise interference problems from the surroundings [26]. These
interfaces need fixed infrastructure and are not wearable in nature, thus limiting its scope. Due to these
limitations, this type of finger motion sensing is not feasible for our application.
2.1.3 Microsoft Digits – Wrist Worn IR Based Finger Sensing
A common problem with all of the above systems discussed is that these systems are not portable or must
completely enclose the hands or fingers inside a glove. Due to their design limitations, these systems are
not present all the time receiving user commands and hence they cannot be called always-on and always-
available interfaces. This fact makes their usage cumbersome, for reasons like arm fatigue that occurs due
to the need to make the hands visible to the sensors in a pre-determined fashion.
Microsoft Digits is an innovative solution that brings wearability and high fidelity sensing along with the
flexibility of vision based systems in the mobile setting. Its design, working and analysis of limitations has
been presented in [16]. This system brings full hand pose detection and reproduction with high fidelity
sensing without the need for wearing gloves. This system consists of a wrist mounted IR camera, IR diffuse
LED, IR laser line projector and an IMU. Two methods of IR illumination are used for faithful finger pose
reconstruction. The camera is placed so that the upper part of the palm and fingers are imaged as they bend
inwards towards the device. A forward kinematic (FK) algorithm reconstructs the 3D pose of the fingers
by exploiting biomechanical constraints on the reflected IR lights from each finger. A complementary
scheme uses a ring of IR LEDs and is used to sense the fingertips. These two sensing methods are combined
to derive an inverse kinematic (IK) algorithm based full reconstruction of the hand and finger pose [16].
However, the hardware is complex as well as costly and the underlying software pipelines are too
sophisticated. Also this being a vision based system, though body mounted, inherits all the shortcomings
of similar vision based systems like line-of-sight sensing and occlusion, that is, the fingers should point
towards the camera for sensing its movements [16].
10
As illustrated by the breadth of the above discussion, about the plethora of available technologies in this
area, it can be inferred that this is a rich and highly challenging research area. However, for designing an
eyes-free, always-on and always-available interface, muscle machine interfaces have generated a lot of
interest. Given their light weight sensing and compact form factor, the capability to capture movements at
various muscle sites and their benefits of having no limitations like occlusion or line-of-sight sensing, they
are an obvious choice for an unobtrusive and wearable interface design. Though their scope of supported
arm gestures or finger movements are practically limited and are often discrete, but for our goal of sensing
a discrete set of finger movements, this interface is an apt choice. The next section presents a discussion on
muscle machine interfaces.
2.2 Muscle Machine Interfaces (MMI)
Muscle machine interfaces are those whose control inputs are signals from our body muscles which are
directly interpreted as user commands by the machine. Muscle-based control has several advantages over
conventional movement-based access devices or technologies. The most prominent among them being no
need for visible physical movement, enabling the user to control the device using only very weak volitional
muscle activity such as twist of the wrist or small finger motions [26-28]. It is also a true eyes-free interface
where the user in command of the machine does not have to pay visual attention to the input interface for
issuing commands. Voice commands are a similar example of such an eyes-free interface, however, voice
commands become too intrusive for others in a public gathering or in a meeting or reveal too much about
user intentions. Muscle machine interfaces are free from such social limitations and awkwardness, and thus
are a perfect fit for an always-on, eyes-free natural user interface. In the following sections we describe
various control signals used for designing muscle machine interfaces.
2.2.1 Surface Electromyography and Its Limitations
Surface Electromyography (sEMG) is the superficial measurement of electrical activity of the muscles. It
is the most popular, commonly used, well understood and well researched muscle based control signal used
11
for both kinesiological and muscle-machine interfaces [12]. Researchers at Georgia Tech and Microsoft
Research have recently shown that EMG based interfaces do not need the user to be in some predetermined
posture , implying that it is a major step towards enabling real-world natural user interfaces [26]. Saponas
et al. have also shown that these muscle machine interfaces are distinctly useful in situations when the
user’s primary commanding organs (say our hands) are already busy doing some other task, say holding a
mug. These interfaces can still issue commands that can be sensed by the machine with remarkable
accuracy. These two advantages of muscle machine interfaces over other types of gesture sensing are the
main impetuses behind our choice of interface to be an MMI.
With all its advantages as noted above, sEMG signal has some problems such as the deterioration of signal
quality with perspiration - consider for example perspiration in the wearable scenarios such as ours [29].
The signal is highly sensitive to the location of sensor placement [30], as reported by Mercer et al. that
electrode position drastically influences the measurements. They have also mentioned that an exact
electrode positioning for sEMG that produces best signals for particular motions may not always be
repeatable. EMG can only reveal superficial muscle activity [31]. sEMG signals are also known to have
interference problems from power lines (PLI), and it is also reported that sometimes the Signal-to-Noise
Ratio (SNR) is very low [32] implying the signal may become unusable due to high noise. Sophisticated
signal processing techniques such as Stationary Wavelet Analysis, are needed to filter out these PLI. sEMG
signals are also known to have signal contamination issues with motion artifacts [32].
2.2.2 Mechanomyography as a Better MMI Signal
A relatively less popular and sparsely researched muscle based control signal is called Mechanomyography
(MMG). Mechanomyography is the measurement of low frequency mechanical vibrations that are produced
by the muscle fibers when we make muscle movements. A detailed overview of the science of MMG is
presented in the next section. The focus of this thesis is on an MMG based MMI due to its unique
capabilities and advantages viz-a-viz sEMG. In contrast to the above noted shortcomings of sEMG, by
virtue of its mechanical nature, MMG is known to be less sensitive to electrical interference problems and
12
is inherently immune to signal degradation due to perspiration. It is relatively less sensitive to sensor
placement [12, 33], and is capable of capturing deeper muscle activities compared to sEMG [29]. MMG
instrumentation is simpler as compared to EMG signal acquisition. A detailed study of MMG sensor design
is presented in Posatskiy’s thesis which states that microphone based sensor for MMG signal acquisition
has better immunity to motion artifacts and are less contaminated by these sorts of unwanted motions [34].
The above mentioned benefits of MMG over sEMG prompted us to choose this as our preferred control
signal for designing the interface.
2.3 Science of Surface Mechanomyography (MMG)
Muscle sounds were first discovered in the 1660s, when Francesco Grimaldi observed a rumbling sound as
he placed his thumbs over the opening to his ears and clenched his fists [35]. But until 1948, due to the lack
of accurate instrumentation and signal storing devices, the research on muscle sounds could not gain
popularity. Thereafter, with the advent of piezoelectric contact sensors such as Rochelle salt crystals,
Gordon and Holbourn in 1948 were the first to record muscle sounds [35]. Muscle vibrations were
historically known by the names of Vibromyogram, Acoustic Myogram, Phonomyogram, Sound Myogram
and Accelerometermyogram before the name Mechanomyogram (MMG) was universally accepted in 1995
[13]. In the next section we discuss the bio-mechanics of MMG signal generation from the muscles and
their properties.
2.3.1 Bio-Mechanics of MMG
Surface mechanomyography (MMG) is the measurement and analysis of low frequency muscle vibrations,
thought to be caused by the movement and deformation of muscle fibers, and reflects the contraction
characteristics of individual motor units [29]. The activation of the motor units cause fiber twitching [36]
and the summation of multiple such twitches produces a gross lateral movement [37]. This is followed by
a period of oscillations which is thought to be corresponding to the vibration at the muscle’s resonant
frequency [37]. These vibrations are termed as acoustic signals that we know as Mechanomyograms.
13
MMG signals are also thought to be generated by the lateral displacements of the muscle and the signals
are propagated thereof. During muscle contractions, the acto-myosin filaments slide over the muscles which
determine the shortening of the long axis of the muscle. These changes in muscle length result in changes
in the transverse axis dimensions of the muscle, thus emitting low-frequency vibrations, which can be
measured on the surface of the skin just above the muscle [27].
The precise source of MMG, though elusive, is thought to be one of these three possibilities –
1. Contraction of muscle fibers causing radial thickening of the muscle,
2. Radial asymmetry along the length of the muscles causing uneven muscle contractions
3. Axial twisting of the muscles.
However, Frangioni et al. indicate that the lateral vibrations are the dominant mode [38]. Taking into
account the degrees of freedom for individual muscle fibers and the selectivity of motor unit recruitment
by the nervous system, from a theoretical perspective, the vibrations we consider to be MMG may be a
combination of many complex phenomenon which cannot be attributed or reduced to unidirectional modes
(e.g. only lateral vibration) [12]. Here we see that considerable amount of research still needs to be done to
ascertain the appropriate source of the MMG signal. However it is certain that the MMG signal is generated
by the movements of the muscles. The next section discusses the characteristics of the MMG signal.
2.3.2 Characteristics of MMG
The acoustic frequency contents of these MMG signals are in the range 3.3Hz – 25Hz [39]. Due to the
varied nature of muscles generating these vibrations and their medium of propagation for each user, the
frequencies are reported to be slightly user specific [14]. However, the frequency band remains constant as
reported above. The vibration patterns of muscles have been shown to occur in a plane perpendicular to its
long axis [38]. Frangioni et al. [35] hypothesized that the frequency content of the MMG signal may be
closely related to muscle stiffness. Oster and Jaffe [40] and Barry et al. [14, 41] have documented that
increasing muscle fatigue causes reduction in the MMG amplitude. This is a very important characteristic
14
of MMG and has to be taken into account while analysis is done for MMG data acquired in longer sessions
and for real-life muscle machine interfaces.
The displacements of the skin surface occurring due to the MMG signals generated beneath them are in the
range of approximately 500nm and the sound volume is in the range of 10dB at a distance of 1mm from the
skin surface [42]. It is also argued that due to the mechanical nature of sound propagation, there is
comparatively lesser need to be very specific in sensor placement location and the MMG signals are
reproducible at a slightly varied location unlike sEMG where sensor placement is a key factor towards
acquiring good signals. Silva et al. have also argued that MMG is not affected by changes in skin impedance
and thus immune to contamination from skin perspirations. This makes MMG to be better suited than sEMG
for the design of wearable access technology solutions.
MMG signal amplitude was reported by Barry et al. to be directly related to the length of a particular muscle
and was argued to be independent of the muscle activation [37]. It is also reported that the signal amplitude
is maximum at the middle portion of the muscles and decreases at the edges of the muscle fibers [35]. Thus
the ideal location for sensor placement is on top of the middle portion of each muscle.
All these characterizations of MMG help us in formulating the signal acquisition and signal-conditioning
strategies for MMG. The task is primarily the detection of low frequency vibrations on the surface of the
skin, a gamut of transducers such as piezo-electric contact sensors, accelerometers, electret condenser
microphones and laser distance sensors have been used in various research and design scenarios. For this
thesis, we have used electret condenser microphones (ECM) fixed at the end of an acoustic chamber as our
MMG sensor, the reasoning for which will be shown in Chapter 3. The sensor design, signal acquisition
methods, their comparisons and our choice of ECM modules as the preferred transducer are discussed in
more details in Chapter 3. The next section discusses some scenarios where MMG has been used as a
control signal in an access technology.
15
2.3.3 MMG Control Signal in MMI
Control applications developed using MMG are mostly focused on development of access technologies for
individuals with a disability. MMG based prosthetic hands [12, 43] and various other MMG triggered
control switches such as eyelid movements [44] have been reported to provide good results for design of
muscle machine interfaces. Research has also been done on developing MMG based body movement
classification systems such as grasp-motion classification [12]. Alves et al. in their 2010 article have shown
the design of a system capable of recognizing coarse hand movements and reported a classification of 7 ±
1 hand movements with accuracy of 90 ± 4 %. The system reported in [27] uses eight silicone mounted
MMG sensors [45] placed at eight different locations on the forearm, used to capture eight kinds of coarse
hand movements. It should be noted here that this system uses one sensors per motion to acquire the MMG
data, whereas our system uses lesser number of sensors for data acquisition, giving comparable results.
Zeng et al. 2009 have reported the results of a hand-motion pattern detection system which detects a set of
four coarse hand and wrist movements with an average accuracy of 79 ± 7 % [46].
Grossman et al. [47] have worked on an MMG based system used to detect finger flexion of the human
hands. They have used five accelerometer-microphone pair sensors as described in [45] placed on the
forearm to capture flexion of five fingers of the hand. This sensor uses an accelerometer mounted over a
microphone in a single package and packed in a silicone chamber. They have reported a classification
accuracy of 67% for thumb flexion and an overall classification accuracy of 76% for classification of flexion
movements among all five fingers. It should be noted here that their focus was to detect and differentiate
only one motion among five fingers, whereas the focus of this thesis is on the detection of four motions of
the human thumb. Similar to [27], in the system mentioned above, one sensor per movement has been used
to acquire data, compared to our system where we use less number of sensors. However our system
produces much better results compared to the results reported for these systems.
As described in the Motivation section of Chapter 1, our goal is to develop a low-cost wearable thumb
mounted joystick. Until this point in this chapter, we have justified our choice of the interface to be a muscle
16
machine interface. As a logical next step, in the next section we present a detailed discussion on thumb
physiology and deduce the muscle locations for the appropriate placement of the mmgSensors.
2.4 Thumb Physiology and Sensor Placement
The human thumb has the capability of unique movements compared with other fingers of the hand. The
thumb is opposable and possesses the capability for motion around multiple axes. Its capacity for
performing movements along varied sets of axes allow us to have a better grip, precise control and various
other functions of hand. These movements are made possible by its unique physiology, the bone structure
supporting the thumb and its musculature and these are described in details in the next section.
2.4.1 Thumb Motion as a Joystick Controller
The saddle shaped trapeziometacarpal joint (TMJ) on the thumb, also called carpometacarpal (CMC) joint,
is unique due to its position in the hand and the possible degrees-of-freedom (DOF) of its movement.
Although there is evidence of three degrees of freedom for the CMC joint, but considering active motion
of thumb during any function, various muscular forces restrict the CMC joint of the thumb and render it
capable of rotation about two axes of motion [48]. These two DOF motions of the thumb, captured reliably,
can be translated to function as a computer joystick which necessarily is a two DOF input device.
To understand thumb movement axes, the thumb TMJ biomechanics should be understood. Hollister et al.
showed that the thumb TMJ is predominantly a two-DOF joint [49]. The two dominant axes of motion of
the thumb TMJ can be modeled as TMJ-Roll (TMJR) axis and TMJ-Pitch (TMJP) axis. These two axes
allow motion in the Flexion-Extension (FE) and Abduction-Adduction (AA) planes respectively. The
position of TMJR is predefined, but that of TMJP is dependent on the thumb abduction and rotation. Though
there is a third axis of motion which facilitates motion in Pronation-Supination (PS) planes, but the motion
in this plane is highly restricted and for a given specific position in AA and FE planes, there is a fixed
position in the PS plane [48, 49].
17
Experimental data, however, show that these two axes of thumb motion are non-intersecting and non-
orthogonal to each other [50]. This, in addition to the variation in the patterns of thumb motion pose the
two most significant challenges for reliably measuring thumb motion. However, the goal being
interpretation of thumb motion as a computer joystick, we need to find four directions where the thumb has
maximum stable measurable movement and the AA and FE planes provide those four distinct directions
for thumb movements.
2.4.2 Thumb Muscles and Sensor Placement
In clinical terms, the thumb is termed as pollex [51] and most of the muscles related to thumb movement
have a term -pollicis in its name. Short muscles responsible for the thumb movements are named thenar
muscles and together they form the thenar eminence. Major thumb motions can be classified into six types
viz. abduction, adduction, flexion, extension, circumduction and opposition, though the last two motions
are composite ones, meaning they are a combination of other kinds of motions. Thumb movements are
controlled by two types of muscles – longus muscles are the long muscles originating in the forearm and
brevis muscles are short muscles originating in the palm of the hand.
The major thumb muscles are as follows–
Abductor Pollicis Longus – It is an extrinsic muscle on the forearm which helps in abducting the
thumb around the carpometacarpal joint, which moves the thumb anteriorly [52].
Flexor Pollicis Longus – It is an obliquely placed muscle on the forearm which assists in flexing
the distal phalanx of the thumb around the inter-phalangeal joint, towards the palm of the hand
[52].
Extensor Pollicis Longus – It is a dorsally placed skeletal muscle that extends the distal phalanx
and the proximal phalanx of thumb [52]. This muscle is needed for hyperextension of the thumb at
the inter-phalangeal joint.
18
Extensor Pollicis Brevis – It is also a dorsally placed skeletal muscle that extends the proximal
phalanx and 1st metacarpal of thumb at the carpometacarpal joint [53].
Abductor pollicis brevis – It is a short muscle of the thenar eminence and abducts the thumb
perpendicular to the palm acting across the carpometacarpal and metacarpophalangeal joints. It also
participates to flex the thumb [53].
Flexor pollicis brevis – It is a short muscle of the thenar eminence that has a deep and a superficial
part. It flexes the thumb around metacarpophalangeal joint [53].
Adductor pollicis – It has two heads namely oblique and transverse. This muscle adducts the thumb
by bringing in the palmar plane. It also helps in bringing the thumb next to index finger.
Opponens Pollicis – It is a small triangular muscle at the edge of thenar eminence which helps in
thumb opposition.
The two snapshots below in Fig 2.1 show the sensor placement on the Flexor Pollicis Longus muscle on
the front of forearm and in Fig 2.2 show Extensor Pollicis Longus muscle on the back of forearm.
Figure 2.1 Snapshot of Sensor Placement over Flexor Pollicis Longus muscle.
19
In order to summarize, we have tabulated the muscles responsible for each of the thumb motions at
carpometacarpal (CMC) joint and present it as follows.
Movement Movement Verbose Muscles Responsible Position in Forearm
Flexion Bending the joint resulting in a
decrease of angle; moving the
bone below the thumb toward
the hand and slightly Forward.
Flexor Pollicis Longus Front of Forearm
Flexor Pollicis Brevis Thenar Eminence
Opponens pollicis Edge of Thenar Eminence
Extension Straightening the joint resulting
in an increase of angle; moving
the bone below the thumb away
from the hand and slightly back.
Extensor Pollicis Longus Posterior of Forearm
Extensor Pollicis Brevis Round Edge of Forearm
Abductor Pollicis Longus Posterior of Forearm
Abduction Lateral movement away from
the midline of the body; moving
the bone below the thumb
toward the front of the wrist.
Abductor Pollicis Longus Posterior of Forearm
Abductor Pollicis Brevis Thenar Eminence
Adduction Medial movement toward the
midline of the body; moving the
bone below the thumb toward
the back of the wrist.
Adductor Pollicis Between Thumb and
Index Finger
Extensor Pollicis Longus Posterior of Forearm
Flexor Pollicis Longus Front of Forearm
Table 2.1 Tabular representation of Thumb Movements and Its Muscles
20
Figure 2.2 Snapshot of Sensor Placement over Extensor Pollicis Longus muscle.
As we see from the table above, the major four motions of the thumb have their related major muscles on
the posterior side and frontal side of forearm. The exact location of these muscles can be palpated on their
specific muscle sites on the forearm while performing respective thumb motions. As described earlier in
this chapter, MMG signals are not highly dependent on muscle sites, thus, placing one mmgSensor each on
the frontal and posterior sides of forearm should provide sufficient MMG data for our task of capturing four
sets of thumb gesture recognition.
An appropriate way of attaching two mmgSensors on the forearm is to design a wristband with the two
mmgSensors embedded in it. A loose wristband will not capture MMG data because an air chamber formed
under the sensor will not be an enclosed chamber to propagate the air pressure to the sensors and higher
pressure on the muscles while acquiring data can pollute the MMG signal. So this consideration should be
taken care while designing the wristband. The wristband should have an adjustable length Velcro band
attached to it to maintain necessary and sufficient pressure on the forearm muscles by the mmgSensor which
should be configurable per user. The sensor design and various trade-offs associated with it are described
in Chapter 3.
21
2.5 Conclusions
In this chapter we have shown that conventional methods employed for gesture or pose recognition of
coarse hand or body movements are inadequate for recognition of fine movements of the fingers of human
hand. We have hypothesized that muscle machine interfaces (MMI) are highly suitable for wearable, eyes-
free, always-on and always-available interfaces. We have compared a few short-comings of a hugely
popular control signal for MMI namely surface-Electromyography (sEMG) with another control signal for
MMI namely Mechanomyography (MMG) and found that MMG is better suited for wearable access
technology scenarios. Finally, we have presented a brief discussion on the suitability of the thumb motion
along the AA and FE axes to be reproduced as a computer joystick. We have also deduced the appropriate
sensor placement locations on the forearm that are suitable for MMG data acquisition for the targeted four
thumb movements. In the following chapters, we will summarize the sensor design and data analysis
methods used in this thesis.
22
Chapter 3 - System Overview and Sensor Design
We have described how physiological signals such as Electromyography (EMG), Electroencephalography
(EEG), and Mechanomyography (MMG) have often been used to design Man-Machine Interfaces and
alternative access devices in Chapter 2. We have also presented a brief description of various methods of
gesture recognition and tried to justify muscle machine interface to be the most suitable solution for the
problem statement targeted in this thesis. A gesture recognition system typically has three subsystems in
common namely, sensors for transducing natural human gestures; data acquisition and signal processing
subsystem; data analysis and pattern recognition subsystem for decisions making. In this chapter we present
an overview of such functional subsystems for the current thesis, their individual functioning, their
interworking and a general data-flow path through the system. Further we also discuss the sensor design in
details with various design decisions made while choosing the transducer for the MMG sensor and finally
describe the data acquisition subsystem.
3.1 System Overview
We have conceptually divided the process of mechanomyography based thumb gesture recognition
applicable to this thesis into four subsystems. The division among these subsystems is done such that the
dataflow between these subsystems are unidirectional and no known interdependencies exist among these
subsystems. Practically, they can be used to form a process pipeline that can be run on a time-overlapped
basis to achieve near real-time finger gesture recognition. However, the development of such a time-
overlapped pipeline is outside the scope of this thesis. This work consists of individual working of each
subsystem and their integration into one completely automated gesture recognition system is in the scope
for future work. The conceptual subsystems as discussed above are presented in the following Table 3-1.
23
Functional Subsystems
Sensor Interfacing and Data Acquisition Automatic Event Annotation
Feature Extraction and Feature Selection Pattern Recognition on Gestural Data
Table 3.1 Functional Subsystems in which the current work is divided.
We present a visualization of the above functional subsystems in the following block diagram in Figure -
3.1 where we summarize the individual subsystem workings. We also give an overview of the complete
system and try to show that there is practically no data inter-dependency among the subsystems and the
data flow is unidirectional, implying the feasibility of future implementation of the process pipeline that we
mentioned earlier.
Figure 3.1 Flow Diagram of Major Subsystems
24
Figure 3.2 System Block Diagram of Major Subsystems
In the following sections we provide a brief description of each of the subsystems here. We will deal with
each of these subsystems in detail in later sections and chapters.
Data Acquisition using Wearable Sensor
The transducer used for data acquisition is a wearable wrist-band with two mmgSensors embedded in it.
These mmgSensors are placed on the forearm of the user at the sites of the muscles responsible for thumb’s
movements, as discussed in Section 2-4. Each mmgSensor comprises of an electret condenser microphone
(ECM) cartridge placed inside an acoustically sealed chamber. The sensor design section of this chapter
presents a detailed discussion on the justification of choice of the transducer. The sound files are saved in
.wav format. The data acquisition for the MMG data is further described in Section 3-3.
Automatic Event Annotation (AEA) Algorithm
Each time series file containing the complete movement data for the current sample is imported in
MATLAB. A measure of energy of the wave is computed by squaring the amplitude. An event start is
25
declared when the total energy of the MMG signal, in an experimental window length is above an energy
threshold, that is set at a value which suits the observations in the samples collected. That is to say that the
threshold values are dataset dependent which implies that these threshold values can be learnt in real time
using various feature learning algorithms and can be thought of as the scope for future work. The automatic
event annotation algorithm is described in more details in Section 4-1.
Feature Extraction and Feature Selection
Each of the samples corresponding to a single finger movement which are annotated by the automatic event
annotation (AEA) algorithm described above, are then used to extract several time-domain and frequency-
domain features. A set of features such as waveform length, zero crossing, slope-sign change, 7th order
autoregressive coefficients, Wilson’s amplitude, mean power frequency, median power frequency of the
signal are used in this project. This feature set represents the major characteristics of each data sample.
Improvements to the current work is possible by searching for more expressive and better features more
suited for MMG signal analysis. A total of sixteen features per channel with two channels of data are
extracted, making it a thirty-two dimensional feature space. This high-dimensional feature space poses a
problem called the “Curse of Dimensionality” which is described in details in Section 4-3. Dimensionality
reduction methods were used which reduces the feature-space dimensionality retaining at least 95% of the
information represented by them.
Pattern Recognition on Gestural Data
We have modeled the gesture recognition for this project as a supervised learning process, which for a
discretely labeled data is called classification of patterns. The dimension reduced data taken from the
previous stage is fed into the classifier algorithms. Two types of pattern recognition algorithms viz.
Quadratic Discriminant Analysis and k-Nearest Neighbor algorithms were employed for data analysis in
this thesis. The theory and working of these algorithms is described in details in Section 4-4. The
classification accuracies on user data is explained in details in the results section which is presented in
Chapter 5.
26
Until now in this chapter, we have briefly presented an overview of the functional subsystems present in
this work. These subsystems will be dealt-with in details in individual sections. The sensor design and data
acquisition subsystem is described in details in the next section.
3.2 mmgSensor Design
As described in Chapter 2, MMG signals are low frequency vibrations emanating from muscles responsible
for the particular motion. These signals have a spectrum of 3.3Hz-25Hz. Mechanomyography is practically
a low-frequency vibration detection problem and many different types of transducers have been used to
capture these muscle sounds. The wearable nature of the problem statement for this project demands
measurement to be done for dynamic muscle contractions i.e. during the presence of motion artifacts. In
the following section we present a discussion on our choice of transducers.
3.2.1 Transducer Selection
There are two methods for capturing Mechanomyography. One is detection of sound or air pressure
vibrations at muscle surface using transducers like accelerometers and the other is detection of muscle
deformation using transducers like laser displacement sensor. Apart from these, various other types of
transducers have also been used for capturing muscle deformation signals such as silicone embedded
sensors [45], piezoelectric contact sensors and ultrasonic sensors [12, 14]. Numerous factors determine the
preference of one over the other such as signal-to-noise-ratio, influence of motion artifacts, sensitivity and
frequency response over the bandwidth of MMG signals. Total cost of the sensor is also a critical parameter
in decision making for this project, because the initial goal of the project was to design a low-cost wearable
interface.
Laser Displacement Sensors
Laser displacement sensors have been used successfully in measuring muscle contractions. Watakabe et al.
has compared these measurements with conventional accelerometer based methods and concluded that the
laser based method to be a standard for MMG measurements. These sensors typically have high bandwidth
27
and high resolution of the order of 5µm, making them an extremely accurate measurement method, but in
addition an expensive and cumbersome one. The necessity for a controlled environment while taking
measurement using laser distance sensors makes them an extremely unsuitable choice for wearable
applications [12].
Piezoelectric Contact Sensors and Accelerometers
Piezoelectric contact sensors and accelerometers are attractive transducers for low frequency vibration
measurements due to their high sensitivity to low frequency vibrations and light weight (less than 3g) [12].
They have the advantage of being mounted directly to the vibrating surface, thus receiving good
measurements. A down side of using accelerometers is that its orientation with respect to the vibrating
surface affects the measurements. Accelerometer measurements also need to be corrected for the offsets
introduced due to using gravity as reference, more so when the vibrating surface is a part of a free moving
object, making it more prone to motion artifacts [34]. Accelerometers also have high cross-axis sensitivity
which makes them highly prone to orientation specificity because positional tilt can affect its measurements.
These drawbacks make accelerometers more prone to noisy measurements, thus making it a less preferable
for dynamic MMG measurements [14].
Electret Condenser Microphone
Pressure based sensing using electret condenser microphones (ECM), which measure the pressure
variations inside an acoustic chamber occurring due to skin displacements caused by muscle vibrations, are
a safer choice in wearable access technology scenarios. Watakabe has shown that the effect of motion
artifacts on the MMG measurements is considerably reduced when using a condenser microphone placed
inside an acoustic chamber [34]. Typically a cylindrical chamber is used, and a condenser microphone is
mounted on the top end of the chamber. The bottom end of the chamber is usually open and pressed against
the muscle site. The transducer may be attached to the skin with an adhesive or placed under a tightened
strap fixed using Velcro [12]. Watakabe has shown that the double integral of the signal recorded using
accelerometer is similar to the signal recorded using microphone, drawing a conclusion that the microphone
28
records skin displacements [34]. These arguments show that for a wearable access technology scenario,
ECM based sensors are best suited due to their immunity to motion artifacts, ease of interfacing, low cost
and robust and simpler designs. In the following section, we will discuss some of the nuances of an ECM
based MMG sensor design, the criticality of the acoustic chamber design and its technical properties.
3.2.2 Design Details of ECM based mmgSensor
Posatskiy et al. (2011) suggest that an MMG sensor should have the following characteristics:
Minimal signal attenuation and a flat frequency response in the range 5Hz–200 Hz.
Maximal attenuation at all other frequencies.
Mass of 5 g or less to avoid muscle activity suppression.
The frequency range of MMG signals falls in the infrasound range 3.3Hz – 25Hz, thus it is very important
that highly sensitive microphones are selected. Professional infrasound microphones used to detect very
low frequency sounds such as Microtech Gefell MK250 have appropriate specifications for infra-sound
measurement, i.e., frequency range of 3.5Hz – 20Khz and sensitivity of 50mV/Pa. However, their
prohibitive cost, to the tune of $2000 per capsule and $1500 per preamplifier, makes them extremely
unsuitable for our purpose. Electret condenser microphones (ECM) are a natural choice due to their better
low-frequency response, low capsule weight and very low cost compared to the professional infra-sound
microphones.
A Panasonic WM-64K microphone was selected here which has a tested frequency range of 20Hz-16kHz
and a sensitivity of 5.6mV/Pa (rated at 1kHz); no manufacturer data is available for inherent noise. The
reason this microphone was selected despite having a cut-off frequency of 20Hz, higher than MMG signal
lowest frequency, is that this figure provided by the manufacturer is their tested value for most ECM
applications. It has been shown that with appropriate preamplifier, these ECMs can work much below the
rated 20Hz frequency [14]. In addition to this, the cost of this ECM is as low as $2 per cartridge and these
are easily available. The preamplifier requirement for WM-64K is far more common - a low noise
29
instrumentation amplifier, as compared to specialized infrasound ECM modules. All these information
reinforces our design choice of using an ECM based MMG sensor for our project.
Using a microphone (ECM) for MMG data acquisition requires custom made acoustic chambers. Chamber
dimensions, as shown by Watakabe in [34], have significant effect on the cut off frequency response and
thus on the signal acquisition. Silva et al. have used silicone embedded microphones for measurement of
MMG signals, where they have placed a thin silicone sheet between the end of acoustic chamber and the
skin-contact [45]. Posatskiy et al. in [12] have shown that using a porous material like silicone can cause
the pressure energy to dampen, thus hampering the measurement. In addition, the dimensions of the acoustic
chamber reported for maximum signal-to-noise ratio (SNR) by Watakabe [34] and Silva et al. [43] have
major contradictions. This is addressed by Posatskiy in his thesis, where he has presented a detailed study
on the characteristics of the MMG sensor viz.-a-viz. the acoustic chamber dimensions. We have taken the
chamber dimensions from Posatskiy’s thesis for our mmgSensor design. In the next section we present the
considerations for the chamber design required for our mmgSensor.
3.2.3 Chamber Design
Posatskiy et al. in his thesis [12] compared chambers of cylindrical and conical shapes of varying chamber
height and diameter and concluded that the conical chamber geometry gives an average increase in signal
gain of almost 6 dB/Hz over that achievable by the cylindrical chambers. This finding challenges the
traditional view and establishes that chamber geometries, despite being significantly smaller than the
wavelength of the signal being measured, can intensify the signal. It also establishes that the chamber
preferably should be made of solid construction to avoid undesirable loss of sound-pressure energy to the
chamber walls such as in case of a silicone housing as used by [43]. This work also focuses on analytically
finding the best conical dimensions for highest signal gain in both cylindrical and conical chamber
geometries. According to Posatskiy et al., the best performing conical chamber dimensions based on mean
gain, low-frequency standard deviation and spectral flatness was in the vicinity of a conical chamber having
a Diameter of 7mm and a Height of 5mm [12, 33].
30
Taking a cue from these findings, we modeled a conical chamber of base diameter of 10mm; top diameter
of 5mm and height of 8mm using the open source 3D modeling software OpenSCAD. We used a
MakerBot3G to 3D print our chamber design. Please note that the dimensions are not exactly what Posatskiy
reported due to some limitations of the 3D printing tool used here. The OpenSCAD model of the chamber
is shown below. Fig 3.3(a) shows the side view, Fig 3.3 (b) shows the top view and Fig 3.3(c) shows the
angular view of the 3D model.
Figure 3.3 Side view, Top view and Angular view of the sensor respectively.
a b
c
31
The above model shows a rectangle bounding the chamber at the top of it, which is meant to be the housing
for attaching the Velcro armband to hold the mmgSensor in place. The armband has adjustable length for
maintaining the correct pressure while tying the sensor to the user’s hand. A correct amount of pressure is
critical, as discussed in Section 2-4, for better signal acquisition because it has been found that higher sensor
pressure over the muscle can adversely affect the signal quality. A snapshot of the sensor embedded in the
armband is shown below in Fig 3.4.
Figure 3.4 Snapshot of the Wristband with two embedded mmgSensor.
An internal line-diagram of the mmgSensor is shown in Fig 3.5 below.
32
Figure 3.5 Line Diagram of the Sensor Housing
Each mmgSensor has a microphone cartridge which is fixed on top of the conical chamber. It should be
noted here that the diameter for the top opening of the conical section is 5mm which is smaller than the
microphone cartridge diameter of 6mm, providing it a physical stopper from moving out of the conical
chamber top. The top of the chamber is acoustically sealed using a hot melt adhesive, which also doubles
as the holder for the cartridge from falling inside the acoustic chamber.
3.3 Sensor Interfacing and Data Acquisition
As described in Section 2-3, we place two mmgSensors on either side of the forearm close to the wrist. This
allows us to capture MMG signals originating from the four thumb movements, namely flexion, extension,
abduction and adduction. The microphone cartridge used for this project is Panasonic WM-64K microphone
as discussed in the previous section of this chapter. The signal is passed through a band pass filter with cut-
33
off frequencies of 2Hz – 50Hz for the pass band. This band of frequencies was chosen as it is suitable for
the MMG signals whose spectrum has been reported to fall in this range.
MMG signals captured using an electret condenser microphone are basically audio signals and this data
acquisition procedure can be treated as a sound recording. Thus, we can use a commercial sound card to
capture the audio signal and store it in a computer hard-drive. We feed the filtered signal from the band-
pass filter into the Edirol FA-101 Audio Interface for pre-amplification and sampling at 8kHz. The filtered
and amplified signal is then recorded in the .wav file format using Audacity open-source audio editing and
recording software. A flow diagram of the data-acquisition procedure is presented in the below Fig 3.6.
Figure 3.6 Flow Diagram of the Data Acquisition Process
The .wav file is the raw audio data file, which is imported in MATLAB for further data processing and
pattern recognition. This includes slicing the time-series data into individual thumb movement events, done
by a custom auto-event sampling algorithm, followed by feature extraction, dimension reduction and finally
pattern recognition. We have described this complete pipeline for data analysis in details in Chapter 4.
3.4 Conclusion
In this chapter, we have presented a brief overview of the entire system including its individual functional
subsystems. These subsystems have unidirectional dataflow among them and thus can be used to build a
process pipeline. We then discuss the sensor design and signal acquisition subsystem in details. We have
34
justified our design choice of using an ECM based MMG sensor by presenting its advantages over other
low-frequency vibration sensing transducers. The acoustic chamber design for the mmgSensor plays a
critical role in determining the sensor characteristics and we have dealt with it in minute details in this
chapter. Finally, we have outlined the sensor interfacing and data acquisition method used for this project.
35
Chapter 4 - Data Analysis and Pattern Recognition
A study of discernible patterns in raw MMG data acquired during finger movement is the most critical step
towards the development of a reliable muscle-machine interface. As discussed in Chapter 2, MMG provides
information about the muscle firings, so patterns of muscle movements will be reflected in MMG data [28].
The raw MMG datasets are treated as a continuous time series and statistical pattern recognition approach
for time series data analysis are applied to find patterns in MMG data. For better automation of the system,
a signal-energy based auto-event annotation algorithm has been developed to demarcate the event
boundaries in the time series as described in the next section. The pattern recognition pipeline used in this
thesis, as described in this chapter consists of data-acquisition; auto-event annotation from the time-series;
representing samples as feature vectors; pre-selection using individual feature discriminatory score called
Fisher’s Score; feature reduction using Principal Components Analysis; and evaluating the reduced feature
set with statistical classification algorithms.
4.1 Auto-Event Annotation (AEA) Algorithm
MMG data is captured as a continuous sequence of finger movements from the two mmgSensors placed on
the forearm. The system captures each movement which produces a fixed duration of muscle sound, and
waits for the next movement thereafter. An example of a raw MMG data capture is shown in Fig 4.1.
36
Figure 4.1 MMG capture for four different kinds of thumb motions
This continuous time-series data needs to be annotated into individual movement samples for individual
feature extraction and pattern recognition. A signal-energy based auto-event annotation algorithm is
presented here. It should be noted that this algorithm assumes prior knowledge of the movement window
and various signal amplitude thresholds. These thresholds are highly data dependent and hence can also be
learned from user data by employing various machine learning algorithms, but the implementation of such
a scheme is out of scope of this thesis. The specific thresholds that are assumed as prior knowledge for this
thesis are shown in the below Table 4.1 and the stepwise algorithm has been presented in Fig. 4.1.1.
Prior Knowledge
Movement window length (mw) Inter-sample window length (isw)
Pre-peak sample length (ppsl) Min amplitude threshold (mat)
Table 4.1 Table listing the parameters assumed to be prior knowledge.
37
Algorithm 4-1 Steps for Auto Event Annotation Algorithm
Auto-Event Annotation Algorithm
1. Compute signal energy (se) by squaring the signal when the value exceeds mat for all data
channels.
2. Compute aggregated window energy (awe) for the time series in the window length mw by summing
se at each sample point for all data channels. Normalize to the maximum awe value.
3. Annotate local energy peaks (lep) by determining the max values of awe in each locality for all
data channels. Please note that there can be multiple such localities in one movement event.
4. Annotate global energy peaks (gep) by determining the max values among lep values in the mw
window lengths for all data channels. Reject all other peaks in isw window length.
5. Annotate movement event begin (meb) by determining the event that is earlier or has significantly
higher amplitude among the channels.
6. Slice the time-series starting from each time = (meb minus ppsl) upto a length of mw. This creates
individual samples for each finger movement.
For better understanding for the algorithm shown above, we are presenting a diagrammatic description of
the steps mentioned in the algorithm using actual MMG sample user data from User-D. Their respective
plots for each steps of the algorithm is drawn using MATLAB and is presented below. Please note that for
all the plots shown below, Time (in seconds) is in the x-axis and Sound Amplitude (5V scaled to 1) in the
y-axis.
Step 1 – The signal values are squared to compute se at each point in time-series. Plot shown in Figure 4.2.
38
Figure 4.2 MMG Signal (Blue) and its Energy (Green) Plots for Channel 1
Step 2 – Compute the aggregate window energy awe from the signal energy se in defined window of
length mw. Plot shown in Figure 4.3.
Figure 4.3 Aggregate Window Energy (Red) Plot overlaid for Channel 1
Step 3 – Annotate local energy peaks lep in each locality which crosses a certain threshold. Plot shown in
Figure 4.4.
39
Figure 4.4 Local Energy Peak (Red) Plots for Channel 1
Step 4 – Annotate the global energy peaks gep which is the largest peak among these lep in a given locality
defined by the mw and ignore every other peak in the isw. Please note that there will be one event annotation
per channel. The next step decides the most appropriate channel, choosing which as starting point will
preserve data from all channels. Plot shown in Figure 4.5.
Figure 4.5 Global Energy Peaks (Blue dots) for Channel 1
40
Step 5 – Annotate movement event begin meb by choosing the appropriate channel whose event begins
earlier or whose amplitude is significantly higher than the others. Plot shown in Figure 4.6.
Figure 4.6 Event Selection among channels (Blue Dots)
Please note that in this case with sample from User-D, the channel having higher amplitudes also has its
event start earlier. So the choice of the channel that decides the governing channel for slicing the time-
series does not change, in this case it is Channel-1. But it can so happen that for a sample from another user,
the channel having earlier event and having higher signal energy amplitude may be different. In such a case,
the channel having the earlier event will be selected as the meb.
Step 6 - Slice the time-series starting from each time = (meb minus ppsl) upto a length of mw. This will
create individual sample files for each finger movement which will be used further for feature extraction
and pattern recognition. Plot shown in Figure 4.7.
41
Figure 4.7 Final Event Demarcation for each Movement.
The variable ppsl accounts for the observation that the transducer recorded event starts a certain amount of
time before the energy peak for the sample occurs. The time-series splits or the event demarcations for each
thumb movement in this sample have been shown with the dark-blue vertical lines in Figure 4.7.
In this section we have shown that the Auto-Event Sampling Algorithm developed for this thesis is able to
annotate and slice the time-series correctly. To the best our knowledge, we are the first group to present
and successfully use such an event annotation algorithm for demarcating and slicing the time-series data.
Other methods, such as additional sensors attached to the moving part to detect the onset of an event, have
been used by other researchers. However, that needs more sensors and hence increases system development
cost. Thus, on developing and successfully using this algorithm enables us to reduce the overall system cost
and also gives us an advantage of easy automation of the data-acquisition. The sliced events from the time-
series enables us to extract features for individual thumb movement class and run pattern recognition
algorithms on them. A detailed description of the feature extraction step has been presented in the next
section.
42
4.2 Sample Feature Extraction
Prudent selection of discriminatory features is highly critical to a successful pattern recognition framework.
Feature extraction refers to the process of finding a set of varied statistical properties or values that can
define the dataset in its entirety. Set of features vary depending upon the properties of the dataset to be
modelled for the particular application, which implies, choice of features to be considered in a system is
highly dependent on the dataset and problem statement. For MMG based gesture recognition, we have
measured a set of 16 time-domain and frequency-domain features from each of the two channels of captured
MMG data. The high feature dimensionality increases the possibility of sample class discrimination. The
subsets of features considered here have shown promising results in classification based on MMG, EMG
and other muscular and physiological signals [27, 28, 46, 54].
4.2.1 Feature-Set Description
For a signal of length N and amplitude 𝑥𝑛 in the nth segment, the features are described as:
Waveform Length (wl) – It is defined as the cumulative length of the waveform over the defined
time segment. This feature is a combined measure of the waveform amplitude, frequency and
duration [28]. It is expressed as
𝑊𝐿 = ∑|𝑥𝑛+1 − 𝑥𝑛|
𝑁−1
𝑛=1
Zero Crossing (zc) – As the name suggests, it is the number of times the signal crosses the zero of
y-axis. Various thresholds can be used instead of using zero, to nullify the effects of source noise.
It represents an approximation of the frequency domain properties of the signal [46, 54]. It is
expressed as
𝑍𝐶 = ∑[𝑠𝑔𝑛(−𝑥𝑛+1 × 𝑥𝑛) × 𝐼(|𝑥𝑛+1 − 𝑥𝑛| > 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑)]
𝑁−1
𝑛=1
43
𝑤ℎ𝑒𝑟𝑒 𝑠𝑔𝑛(𝑥) = {1, 𝑥 > 00, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝑎𝑛𝑑 𝐼(𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛) = {1, 𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛 𝑖𝑠 𝑡𝑟𝑢𝑒 0, 𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛 𝑖𝑠 𝑓𝑎𝑙𝑠𝑒
Slope Sign Change (ssc) – Similar to ZC, this feature also approximates the frequency domain
properties. It is calculated by counting the number of times the slope of the signal waveform
changes from positive to negative or vice-versa [46, 54]. A threshold value can be used to nullify
the effect of noise. It is expressed as
𝑆𝑆𝐶 = ∑ 𝑓((𝑥𝑛 − 𝑥𝑛+1) × (𝑥𝑛 − 𝑥𝑛−1))
𝑁−1
𝑛=1
𝑤ℎ𝑒𝑟𝑒 𝑓(𝑥) = {1, 𝑥 > 00, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
Wilson’s Amplitude (wa) – It is the defined by the number of times the difference in amplitude of
successive segments of MMG is greater than a threshold [27, 54]. It is reported to be related to
muscle contraction level and firing of muscle units. It is expressed as
𝑊𝐴 = ∑ 𝑓(|𝑥𝑛 − 𝑥𝑛+1|)
𝑁−1
𝑛=1
𝑤ℎ𝑒𝑟𝑒 𝑓(𝑥) = {1, 𝑥 > 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
We have taken the threshold value equivalent to be for our experiments 5mV.
Integrated MMG (immg) – It is calculated as the absolute value of the amplitude of the MMG
signal. Generally, it can be used as the onset detector for the movement event [54]. It is expressed
as
𝐼𝑀𝑀𝐺 = ∑|𝑥𝑛|
𝑁
𝑛=1
44
Mean of Absolute Deviation (mad) – This feature measures the nature of atypical observations in
the dataset [55] and is expressed as
𝑀𝐴𝐷 =1
𝑁∑|𝑥𝑛 − 𝑥𝑛̅̅ ̅|
𝑁
𝑛=1
Mean of Absolute Value (mav) – It is similar to average rectified value and a popular feature in
EMG analysis for detecting muscle contraction levels [54]. It is expressed as the average of absolute
value of the MMG signal amplitude
𝑀𝐴𝑉 = 1
𝑁∑|𝑥𝑛|
𝑁
𝑛=1
Root Mean Square Amplitude (rmsa) – Root mean squared amplitude has been used as an indicator
of force and non-fatiguing contraction in muscular analysis [54]. It is expressed as
𝑅𝑀𝑆 = √1
𝑁∑ 𝑥𝑛
2
𝑁
𝑛=1
7th order Auto-regressive Coefficients (ar7) – Autoregressive model estimation coefficients have
been used previously for retrieving major information from a time-series [46]. Autoregressive
model describes each sample of a time-series as a linear combination of previous terms and white
noise. The notation AR(p) refers to autoregressive model of order p and is expressed as
𝐴𝑅7 = − ∑ 𝜑𝑖𝑥𝑡−𝑖
𝑝
𝑖=1
+ 휀𝑡
𝑤ℎ𝑒𝑟𝑒 𝜑𝑖𝑖𝑠 𝑡ℎ𝑒 𝐴𝑅 𝑐𝑜𝑒𝑓𝑓𝑖𝑐𝑖𝑒𝑛𝑡 𝑎𝑛𝑑 휀𝑖 𝑖𝑠 𝑤ℎ𝑖𝑡𝑒 𝑛𝑜𝑖𝑠𝑒.
7th order AR was suggested for pattern recognition of MMG data in previous research by Alves et
al. [27], so we have considered p=7 for our experiments.
45
Mean Frequency (mpf) – Mean frequency is defined as the average frequency of the power
spectrum and is calculated as the sum of product of power spectrum in a given frequency bin and
frequency of that bin divided by total sum of power spectrum [54].
𝑀𝑃𝐹 = ∑ 𝑓𝑖𝑃𝑖
𝑀
𝑖=1
∑ 𝑃𝑖
𝑀
𝑖=1
⁄
where 𝑓𝑖 is the frequency of 𝑖𝑡ℎ bin and 𝑃𝑖 is the power in 𝑖𝑡ℎ frequency bin. M is called the length
of the frequency bin and is usually taken as the next-power-of-2 of the total length of the time-
series.
Median Frequency (mdf) – MDF is a frequency at which the power spectrum is divided into two
regions with equal amplitude [54] and is expressed as
∑ 𝑃𝑖
𝑀𝐷𝐹
𝑖=1
= ∑ 𝑃𝑖
𝑀
𝑖=𝑀𝐷𝐹
=1
2∑ 𝑃𝑖
𝑀
𝑖=1
In this section, we have described a set of sixteen features that are extracted out of each movement sample
for each channel of the user MMG data. There is ample scope for future research in finding a more suitable
feature set for MMG dataset which may improve the overall classification of the system. However, these
features presented above give us a good representation of the MMG dataset used for our research which is
evident from our final classification results presented in Chapter 5. With big feature set and smaller
observation set, we face a problem popularly called curse of dimensionality which is described in the
following sub-section. The remedy for this problem is dimensionality reduction.
4.2.2 Curse of Dimensionality
The basic premise of data analysis is to express the dataset into a rectangular matrix with N rows and D
columns, where rows are different samples or instances and columns are the various properties or attributes
pertaining to each observation. In statistical terms, the dataset is described as one having N observations
46
with D variables or features each. This type of analysis is also termed as D dimensional data analysis
because the data visualization can be thought of as N points in a D dimensional space. In the usual case of
N > D, the statistical learning problem statement is tractable, but in cases where N < D, the problem becomes
intractable [56].
In machine learning and pattern recognition, the sample distance is a critical metric for arriving at decisions.
It can be shown that at higher dimensions, the distance metric loses its meaning resulting in difficulty for
computation of nearest neighbors [57]. This means all observation points in the training class in a high-
dimensional space appears to be almost equidistant from a test point. Hence, we cannot distinguish training
points of one class from another, and thus cannot classify our test points. This poses a problem, popularly
known as the curse of dimensionality.
The term curse of dimensionality, coined by Richard Bellman, suggests that convergence of any estimator
to the true value of any smooth function defined in the higher dimensions is very slow [56]. This means
that total observations, N , needed for the convergence of the solution to the true value within a desired
level of accuracy in a learning problem increases exponentially with increasing dimensionality, D , of the
dataset. One prominent problem of the curse of dimensionality is that, at higher dimensions duplicity of
features increases. There are mathematical methods using which the dimensionality of a dataset can be
reduced without significant loss of information conveyed. One such effective method is Principal
Components Analysis which is an unsupervised dimension reduction method, meaning the dimensionality
reduction is done irrespective of the class discriminatory information present in the data. The Fisher
Discriminant score is another such dimensionality reduction method which is a supervised reduction
method. This method takes the class labels from the data in addition to the raw data matrix and ranks the
features according to the class discriminatory information present in them. Both these methods have been
described in details in the following section.
47
4.3 Dimension Reduction and Feature Selection
As described in Section 4.2, the dimensionality of current two-channel MMG dataset is thirty-two. When
using a dataset of finite (N) observations per class in training set, classification error depends on feature
dimensionality (D), classifier used and asymptotic probability of misclassification (PMC). For Fisher’s
Discriminant Analysis classifier, D < (N/4.5) assuming PMC to be 0.01. It should be noted that more
features may be incorporated assuming higher estimated PMC [27]. For this project we have used 2 ≤ 𝐷 ≤
4, 𝐷 being an integer. For overall dimensionality reduction on the MMG dataset, we have used pre-
selection based on Fisher Discriminant score for ranking various features and applied Principal Components
Analysis for obtaining the final dataset with lower feature count.
4.3.1 Pre-selection using Fisher Discriminant Score Analysis
The Fisher score (𝐽𝑑) is a pairwise ranking score for evaluating the class-discriminating information present
in each of the features in the class label pair and is defined as the ratio of between-class scatter to the average
within-class scatter. Generalizing it to a dataset with 𝐾 classes, we need to find ½*K*(K-1) number of
pairwise Fisher Scores. For a multiclass classification problem, the Fischer Discriminant Score is expressed
by the formula shown below:
𝐽𝑑 = ∑ ∑ 𝑝𝑖𝑝𝑗(𝜇𝑖,𝑑 − 𝜇𝑗,𝑑)2
𝐾
𝑗=𝑖+1
𝐾−1
𝑖=1
(𝑝𝑖𝜎𝑖,𝑑2 − 𝑝𝑗𝜎𝑗,𝑑
2 )−1
where 𝐾 is the number of classes in the dataset, 𝑝𝑖 is the priori probability of class 𝑖, 𝜇𝑖,𝑑 and 𝜎𝑖,𝑑 are the
mean and variance of the 𝑑𝑡ℎ feature for class 𝑖 [27]. For the current MMG dataset, we have four classes of
data labels pertaining to the four thumb movements and thirty-two features for each observation.
4.3.2 Principal Components Analysis (PCA)
PCA is an orthogonal linear transformation of the dataset which is widely used for dimension reduction in
data analysis [58]. Though there are various methods of performing PCA, we have selected the method of
48
PCA using covariance as was suggested by Zeng et al. [46]. By finding the eigen-values of the covariance
matrix, the process finds and aligns the basis (coordinate system) of the dataset towards the direction of
maximum variance, and thus making it easier to discriminate among various classes. Suppose we have a
dataset 𝑋 = {𝑥1, 𝑥2 … 𝑥𝑛} with 𝑁 observations and 𝐷 features or dimensions such that 𝑋 is a rectangular
matrix of 𝐷 rows and 𝑁 columns. The goal of PCA is to find a new dataset 𝑌 = {𝑦1, 𝑦2 … 𝑦𝑛} with 𝐿
features (𝐿 < 𝐷), where this set of 𝐿 features conveys the most relevant information contained in the
original dataset [46]. The steps needed to perform for PCA are described below:
Find the covariance matrix of the mean-corrected dataset – Compute the mean corrected
matrix, 𝐵 which has zero mean for each row. Compute the covariance matrix (CM) for 𝐵 by using
the following expression
𝐶𝑀𝑖𝑗 = 1
𝑁 − 1∑(𝑥𝑖 − 𝑥�̅�)(𝑥𝑗 − 𝑥�̅�)
𝑁
𝑖=1
∀ 𝑖 ≠ 𝑗
Find the Eigenvalue Decomposition of 𝑪𝑴 matrix – Find the eigen-values and corresponding
eigen-vectors of the covariance matrix from previous step. Let 𝑒𝑉𝑎𝑙 = {𝜆1, 𝜆2 … 𝜆𝐷} be the eigen-
values and 𝑒𝑉𝑒𝑐 = {𝑒𝑣1, 𝑒𝑣2 … 𝑒𝑣𝐷} be the corresponding eigen-vectors of the 𝐶𝑀 matrix.
Forming a New Feature Set - From the eigenvalue decomposition of the existing dataset, select
𝐿 largest eigenvalues such that the sum of these eigenvalues correspond to more than a threshold τ,
(say 95%) of the information in the original dataset. Let 𝑒𝑉𝑒𝑐 = {𝑒𝑣1, 𝑒𝑣2 … 𝑒𝑣𝐿} be the chosen
eigenvectors, then the reduced feature dataset will be represented by
𝑌 = 𝑒𝑉𝑒𝑐𝑇 × 𝑋
𝑤ℎ𝑒𝑟𝑒 𝑒𝑉𝑒𝑐𝑇 𝑖𝑠 𝑡ℎ𝑒 𝑡𝑟𝑎𝑛𝑠𝑝𝑜𝑠𝑒 𝑜𝑓 𝑒𝑉𝑒𝑐.
The step of choosing 𝐿 highest valued eigenvectors is the main feature selection step in PCA.
This problem of dimensionality reduction can be thought of as an optimization problem where we have to
select 𝐿 features out of the original 𝐷 features, which must fulfill the constraint of minimum threshold data
49
representation, i.e. more than 95% information of original dataset and must optimize classification
accuracy. Alves et al. has shown the use of genetic algorithms (GA) to solve this optimization problem, but
such approaches are out of scope of this thesis. We search for the best representing set of features by
experimentation. Once we have obtained the reduced feature dataset 𝑌 that represent more than 95% of
original dataset, it can be treated as the original dataset, but with reduced feature size, on which we have to
run pattern recognition algorithms. The next section discusses the pattern recognition algorithms that have
been employed for compiling the classification accuracies for this thesis.
4.4 Pattern Recognition
We have modeled this problem of finger gesture recognition as a supervised learning problem, which for
discrete labeled data is also known as a classification problem. A supervised machine learning problem is
one for which the classifier has the class-labels as an input, i.e., the classes for each observation is known
to the algorithm. This means that our muscle machine interface based on MMG signals will have a training
phase, where the user will train the interface with a pre-specified number of samples for each movement,
before it can be used as an access technology. For experimentation purposes with limited data, we collect
MMG data and divide the dataset into a train and a validation set. Classification rule or classifier algorithm
learns the classification parameters by training on the test subset and further uses these learned parameters
to predict the unknown class of the new data points in test subset.
There are various types of classifier algorithms, that can be used for supervised learning scenarios, such as
multilayer perceptron, deep belief networks, ensemble methods and support vector machines, but most of
these algorithms have higher computational footprint. A near real-time performance is critical for the
problem statement of this thesis. As shown in [27, 46], two efficient and simple implementation based
classifiers are Quadratic Discriminant Analysis (QDA) and k-Nearest Neighbors (kNN), which have been
used for MMG based pattern classification problems. We have used these above mentioned classifier
50
algorithms to compute the final classification accuracies and presented a comparison between them in the
results chapter. The next section presents a discussion of these two algorithms.
4.4.1 Quadratic Discriminant Analysis (QDA)
Consider a new dataset 𝑌 with reduced dimensionality, obtained using the methods mentioned in Section
4.3, as a training set for the QDA classifier. 𝑌 is a set of 𝑁 observation vectors per class having L dimensions
for each observation. Each class of data is pertaining to one type of thumb movement. The method of
classification based on QDA is described below:
Assumptions:
Let 𝑌𝑐 be the class label vector for training set 𝑌;
πk be the prior probability of an observation belonging to a particular class.
Obviously, ∑ 𝜋𝑘𝐾𝑘=1 = 1 where 𝐾 is the number of classes in the dataset;
Assume the covariance matrix to be different for each class, denoted by 𝛴𝑘. Considering covariance
structure to be same for all classes will make the decision boundaries to be linear.
Bayes rule says that the optimal solution for such a problem is to choose a class which maximizes the
posterior probability given the observation vector 𝑌. The Quadratic Discriminant Function is:
𝛿𝑘(𝑦) = −1
2𝑙𝑜𝑔𝛴𝑘 −
1
2(𝑦 − 𝜇𝑘)𝑇𝛴𝑘
−1(𝑦 − 𝜇𝑘) + 𝑙𝑜𝑔𝜋𝑘
The new observation i.e., the observation from the test set will be assigned to a class 𝑘 which maximizes
the discriminant function 𝛿𝑘(𝑦) [59]. In case of this thumb gesture recognition problem, the number of
classes considered is four and prior probability of each class is considered equal i.e., 0.25 each. Assuming
the classes to be equi-probable is a sound assumption because in the real scenario, all the four movement
classes may occur with equal probability, and hence the prior should reflect that.
51
4.4.2 k-Nearest Neighbor (kNN)
The kNN algorithm is a simple yet useful classification algorithm where the classification rule is to measure
the distance of a new 𝐿-dimensional observation point in the test set, from all the training points using any
of the distance metrics such as Euclidean distance. Assign the new observation point to that training class
label which occurs the most number of times among the k shortest distance training points from this test
point.
The kNN algorithm can be summarized as:
Define the distance metric, say Euclidean, on input the vector Y.
Classify new data point to modal class label of the k- nearest points in training set.
Given a particular distance metric, the only parameter to tweak in this algorithm is the number of neighbors,
𝑘, to consider while computing the distances [59]. The choice of 𝑘 is highly data dependent and is also very
critical in determining the complexity of computation. Feature normalization before applying the kNN
algorithm is sometimes done to ensure that the features play equal role in determining the distances, but its
use is rather objective and we have not used it in our project.
In the prior two subsections, we have discussed the classifiers used in this thesis namely, QDA and kNN.
As reported earlier, both these algorithms have been used in prior research using MMG dataset. The results
are discussed in details in the following chapter. In the next section, we discuss the cross validation strategy
used to compile the classification accuracies in this thesis.
4.4.3 Cross Validation of Test Samples
Cross validation is a very useful technique while working with small datasets for ensuring that the designed
system and its algorithms work for independent and unseen datasets. For doing this, the original dataset is
randomly split in training-set on which the classifier is trained and test-set on which the classifier parameters
are verified. The samples are randomly interchanged multiple times between these two sets and the
52
algorithms are run as many times as the count of data points in the training set. The final result of
classification accuracy is obtained by majority vote or other data aggregation methods. This is called cross
validation.
One extreme case of this cross-validation method is called Leave-One-Out-Cross-Validation or LOOCV.
Here, the classifier is trained on (N-1) data points of the training set and the left out one data point is
predicted. This is repeated N times to cover all data points in the training set. Let 𝑌 = {𝑦1, 𝑦2 … 𝑦𝑛} be the
original dataset with N observations. For this dataset, the process of LOOCV will be run 𝑛 times producing
classification accuracy results 𝐶 = {𝑐1, 𝑐2 … 𝑐𝑛}. Then the final classification accuracy result will be the
mean of all these values.
4.5 Conclusion
In this chapter we have presented a detailed overview of the data analysis method used in this thesis. We
started with the description of the feature extraction methods with details on the various time and frequency
domain features used for this MMG dataset. Thereafter, we discussed the curse of dimensionality and its
remedy i.e. the dimensionality reduction methods namely, Fisher score based feature pre-selection and PCA
based dimension reduction. Finally, we discussed the classifier algorithms used to predict the class of the
new test observation points namely, QDA and kNN. For working with datasets of limited size, cross
validation is an important tool. We have discussed an extreme version of this namely LOOCV, in the last
section, which is used to compile the results that are presented in the next chapter.
53
Chapter 5 - User Study and Results
In this chapter we discuss the experiment design, user study design methods and criterion for the choice of
users. We then discuss the metric for comparing our classification results and finally compile and compare
various algorithms for producing final results.
5.1 Experiment and Methods
A user sample of six able-bodied volunteers consisting of both healthy males and females in the age group
22-28 years, provided written consent to participate in the experiment. Participants reported they had intact
wrist musculature and they did not have any history of neuromuscular disorder. These participants are
referred to as Users A-F in this thesis. The apparatus and methods to be employed in this user study has
been approved by Virginia Polytechnic Institute and State University’s Institution Review Board protocol
bearing serial number IRB#13-512.
5.1.1 Experimental Protocol Design
Participants sat on a chair with their arms on the armrest. The participants’ hand was positioned in a way
such that the palm and fingers of the hand were free to move and the MMG sensors placed on the forearm
were not touching the armrest. Maintaining such a posture where the MMG sensor is free from external
mechanical disturbance is important because the MMG measurement using the mmgSensor is based on
capturing audio signals placed at the end of a reverberation air-chamber, and any external movement on the
sensor may hamper the MMG data. The sensors were affixed to the participants’ forearm using a Velcro
armband attached to the custom sensor housing as shown in Chapter 4.
As already described in Section 2.4, the muscles on the forearm suitable for capturing MMG data pertaining
to four thumb movements; abduction (AB), adduction (AD), flexion (FL) and extension (EX) are Flexor
Pollicis Longus (FPL), Abductor Pollicis Longus (APL) and Extensor Pollicis Longus (EPL). Since, for
this project, we had decided to use only two mmgSensors, thus, per Table 2.4.1 the most suitable muscle
54
sites for placing the mmgSensors that are common among most of the thumb motions are FPL (on the front
of forearm) and EPL (on the posterior of forearm). FPL is directly responsible for flexion (FL) and EPL for
extension (EX) of the thumb. In addition, EPL also contributes to thumb adduction (AD). Discarding the
APL muscle site for MMG data acquisition may provide poorer signals for thumb abduction (AB) motions.
This design choice was made in view of an assumption that abduction (AB) is a more pronounced thumb
movement that can produce stronger MMG signals compared to adduction (AD). Thus a more direct access
is required to the muscle site responsible for thumb adduction (AD) compared to thumb abduction (AB).
The sensors were placed on the above mentioned muscles after palpating the forearms while the user
performed the four thumb movements under study, to locate the exact locations of the muscles under study.
The stray wiring for the sensor interface was strapped to the forearm using another Velcro band to reduce
any noise arising from their movements as they were directly connected to the microphones inside the
mmgSensor. On completing the sensor setup, each user was given some time to get accustomed with the
sensor attachment so that they did not feel any sort of extra stress on their muscles while performing thumb
movements which could potentially adversely affect the MMG data. A demonstration of the four thumb
movements was given to each participant by the investigator until they could perfectly perform these
motions. The sequence of thumb movements considered for this experiment was neutral, adduction,
abduction, flexion, extension. The thumb movements are shown in the snapshots below in Fig 5.1 and Fig
5.2. Each subsequent motion was performed at an interval of approximately 1s, allowed to return to neutral
position and then again prompted to perform the next motion. A manual trigger was used for starting the
data acquisition software and voice cues were used by the investigator to approximately time each motion.
The participants were asked to respond only to the voice cues before performing each motion.
55
Figure 5.1 Snapshots of Neutral, Abduction and Adduction positions (L-to-R)
Figure 5.2 Snapshots of Flexion and Extension positions (L-to-R)
To the best of our knowledge, MMG data acquisition on this set of thumb motions has not been attempted
before and we are the first to report an experiment using these four thumb motions. Several users reported
56
thumb muscle fatigue in repeatedly performing thumb abduction and thumb extension. Muscle fatigue has
been reported to have adverse effect on the signal quality of MMG data. Therefore, owing to the fatigue
reported by the users, some samples had to be cut short to reduce the effect of muscle fatigue on data.
5.2 Results
This section is organized into subsections that first describe the metric used for comparing classification
results obtained from two classifier algorithms, the choice of dimensionality that gives best classification
accuracy and finally compare the results from each classifier based on these metrics.
5.2.1 Metric Used for Comparing Results
Average Classification Accuracy is a measure of performance of classifier algorithms which is defined as
the ratio of correct classifications to the total number of samples. This is a popular metric for comparing
the results of pattern recognition algorithms.
In addition, to determine the number of movements that can be reliably classified by the various algorithms,
we formulate a new metric which we name as the True Rate of Classification (TRC) and define it as:
𝑇𝑅𝐶 ≈ ℱ(𝐴𝑣𝑔. 𝑅𝑎𝑡𝑒 𝑜𝑓 𝐶𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛 , 𝑁𝑜. 𝑜𝑓 𝐶𝑙𝑎𝑠𝑠𝑒𝑠 𝑖𝑛 𝐶>75)
where the set 𝐶>75 is defined as the set of movement classes that can be classified with an average accuracy
of more than 75%. A simple function that can be used to compute this metric, is the product of these two
parameters namely the average rate of classification and average number of classes in the set 𝐶>75. Results
compiled using both of these metrics are presented in Table 5-1 through Table 5-3.
5.2.2 Choosing Dimensionality for this MMG Dataset
We need to find the optimal dimensionality, D, for this problem that gives both the highest rate of average
classification and the highest number of correctly classifiable movements for each classifier algorithm.
Thus, the metric defined in the previous section for TRC can be used for determining the optimum
57
dimensionality D such that it gives the highest value for TRC. For this purpose we use QDA algorithm to
compute the average classification rate and the PCA algorithm to reduce the dimensionality of the problem
in the range 2< 𝐷 < 5, (𝐷 ∈ 𝐼) and plot them against each other. The secondary vertical axis in this plot
is the number of classes in the set 𝐶>75 which is shown in Figure 5.3.
Figure 5.3 True Rate of Classification for the entire Dataset for 2<D<5 (D ∈ 𝐼)
From the above definition of TRC and Figure 5.3, clearly we see that for D=4, both the average rate of
classification and the number of classes in the set 𝐶>75 have a peak when compared to other D values. This
implies that for this dataset, QDA algorithm performs best with D=4 and is capable of correctly classifying
3.2 ± 0.75 (or 3 ± 1) types of thumb motions with an average accuracy of 90.5 ± 4.1 % as shown in Table
5-1. Thus going further, evaluations and analyses of the classifier algorithms in this thesis will be done for
a reduced dimensionality of D=4. A detailed description of results obtained from these algorithms is
presented in the following sections.
0
1
2
3
4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
2 3 4 5
Nu
mb
er o
f C
lass
es C
orr
ectl
y C
lass
ifie
d
Cla
ssif
icat
ion
Acc
ura
cy (
%)
PCA Reduced Dimensionality
True Classification Rates for QDA
Avg. Classification Number of Correct Classes
58
5.2.3 Patterns Classification of Thumb Movement Using QDA
Table 5-1 represents the individual classification accuracies of thumb movement data patterns for each
participant evaluated using QDA classifier and Leave One Out Cross Validation (LOOCV) method for an
MMG annotation window length of 600ms. Results in this table are reported as participant specific thumb-
movement pattern classification values on PCA Reduced Feature set with D=4 and with information
retention threshold τ1= 95% that gives maximum pattern classification accuracy.
User
PCA Reduced Feature Set (D=4)
No. of Movements
Avg. Classification Accuracy in % (Mean ± SD)
No. of Movements Discernible at ≥75%
Adjusted Accuracy in % (Mean ± SD)
A 4 80.7 ± 10.6 3 85.2 ± 6.4
B 4 92.5 ± 9.5 4 92.5 ± 9.5
C 4 59.5 ± 6.2 0 NA a
D 4 82.5 ± 13.6 2 95.5 ± 6.4
E 4 91.6 ± 6.8 4 91.6 ± 6.8
F 4 81.5 ± 12.5 3 87.5 ± 0
Avg. (A-F) 4 81.5 ± 12.01 3.2 ± 0.84 90.5 ± 4.1
Table 5.1 Participant-wise Classification Accuracies using PCA and QDA Algorithm a – Left out of calculations for this metric because of the special case with User C presented in Discussions
With this method of data analysis, pattern classification rates were user dependent, being highest for User
E and consistently low for User C irrespective of whichever method was used for feature dimensionality
reduction. As shown in Table 5-1, using the classification accuracy metric, an average pattern classification
of 81 ± 12.01 % was achieved for all users based on a reduced dimensionality feature set of D=4 using only
PCA.
To determine the number of movements that can be reliably classified by the classifier for each participant,
we decided to eliminate those movements whose individual pattern classification accuracy for each user
was below 75% and then compute a new metric called TRC as discussed in Section 5.3.1. On filtering
59
results using this threshold, the classifier could recognize 3.2 ± 0.8 (or 3 ± 1) thumb movements with an
average classification accuracy of 90.5 ± 4 % using PCA based feature reduction only, as shown in Table
5-1.
A different method of feature pre-selection was also employed wherein the features were ranked as per their
respective Fisher Scores which is assigned based on their individual class-discriminatory information. The
criterion for pre-selection of features based on Fisher Score was to choose first N features, which gives best
results on the total classification accuracy metric. This new set of features is run through PCA algorithm to
further reduce the feature dimensionality, thus producing a new set of eigenvectors that are different from
the previous method, where only PCA was applied on all the features.
The results obtained using this method are compiled in Table 5-2 where the best case classification accuracy
for each participant employing the above mentioned feature selection method are noted. This pre-screening
of features based on Fisher Scores followed by PCA for feature reduction improves classification accuracy
for User C whose performance was very poor using only PCA. Employing this method User C shows good
pattern classification rate of 81 ± 6.2 % for all 4 motions, however it should be noted that the dimensionality
had to be further reduced to 2 for User C to achieve this rate of classification. This method also improves
classification accuracy for User E and the classifier can now recognize all 4 motions with an accuracy of
95 ± 8 %. For other users, the overall classification accuracy remains unchanged at previous values that are
obtained using only PCA for dimension reduction. The overall best case classification rate for all the users
(A-F) improves to 85.9 ± 6.9 % as compared to Table 5-1 obtained using only PCA.
True Rate of Classification (TRC) also improved on a feature set chosen using pre-selection by Fisher Score
followed by PCA based dimensionality reduction, as can be seen from the Table 5-2 below, the classifier
could recognize 3.3 ± 0.8 (or 3 ± 1) classes of thumb motion at an average rate of 91 ± 5.9 %, as shown in
Table 5-2. It should also be noted that in using the latter method of feature reduction, all four thumb
movements obtained from User C could be reliably classified, unlike using previous method of feature
reduction.
60
Use
rs
Fish
er S
core
bas
ed F
eatu
re
Dim
ensi
on
alit
y
PC
A b
ased
D
imen
sio
nal
ity
No
. of
Mo
vem
ents
Avg
. C
lass
ific
atio
n
Acc
ura
cy in
%
(Mea
n ±
SD
)
No
. of
Mo
vem
ents
Dis
cern
ible
at ≥
7
5%
Ad
just
ed
A
ccu
racy
in %
(Mea
n ±
SD
)
A 16 4 4 80.7 ± 10.6 3 85.2 ± 6.4
B 16 4 4 92.5 ± 9.5 4 92.5 ± 9.5
C 2 2 4 81.25 ± 6.2 4 81.25 ± 6.2
D 16 4 4 82.5 ± 13.6 2 95.5 ± 6.4
E 4 4 4 95.8 ± 8.3 4 95.8 ± 6.8
F 16 4 4 81.5 ± 12.5 3 87.5 ± 0
Avg. (A-F) 4 85.9 ± 6.5 3.33 ± 0.81 91 ± 5.9
Table 5.2 Participant-wise Classification Accuracies using Fisher Score, PCA and QDA Algorithm.
Finding a common set of optimum features that gives best classification results across all users, employing
just PCA or Fisher Score based pre-selection methods is difficult. It can be seen from Table 5-2 that one
common set of such parameters for PCA or Fisher Score could not be arrived. Other methods of searching
for optimal features in a dataset, such as using Genetic Algorithms as shown by Alves et al. in [27], to arrive
at the best combination of discriminatory features for a particular dataset can also be used, but such an
implementation is out of scope of this thesis.
The most frequently misclassified muscle movement was thumb abduction (AB) with low pattern
classification accuracy. When using the TRC metric, which takes the count of movement classes being
classified with more than 75% accuracy i.e. the set 𝐶>75, thumb abduction (AB) was the movement class
that did not appear in the set 𝐶>75 most number of times. Figure 5.4 shows the classification accuracies for
each user for each thumb movement grouped per-movement. Data from Table 5-2 has been used to generate
these plots.
61
Figure 5.4 Classification Accuracy per-User per-Class employing QDA Algorithm
From Figure 5.4, we see that the spread of classification accuracies for the motion group Thumb-Abduction
(AB) is the most scattered which implies that this motion is the most uncertain for classification. This result
is in accordance with the experiment-design choice described in Section 5.1.1, for selecting two muscles
sites out of three possible options for placing the two mmgSensors. For example, for User-D and User-F,
this motion cannot be classified with a minimum classification accuracy of 75% and thus, it did not appear
in the set 𝐶>75 while computing classification accuracies using the TRC metric.
5.2.4 Patterns Classification of Thumb Movement Using k-NN
Obtaining an optimum k value which produces best classification accuracy for a dataset is highly dataset
dependent, and thus needs to be determined by sweeping across various k values. In Figure 5.5, the average
classification accuracy is on the vertical axis and various k values are listed on the horizontal axis.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A D D U C T I O N A B D U C T I O N F L E X I O N E X T E N S I O N
Cla
ssif
icat
ion
Acc
ura
cy (
%)
Thumb Movement Types
Classwise Class i f icat ion Accuracy per User
User-A User-B User-C User-D User-E User-F
62
Figure 5.5 Classification Accuracies per-User obtained by k-Nearest Neighbor Algorithm
Figure 5.5 represents the individual classification accuracy of thumb movement data patterns for each
participant evaluated using k-Nearest Neighbor Algorithm where 1 < 𝑘 < 10, (𝑘 ∈ 𝐼) and Leave One Out
Cross Validation (LOOCV) method for an MMG annotation window length of 600ms. The dimensionality
of MMG samples are reduced to D=4 using PCA, with information retention threshold τ1= 95% which gives
highest pattern classification accuracy.
To converge on one value of k that gives good results, we find the classification accuracy using the TRC
metric as defined in Section 5.2.1. For using the TRC metric, the average rate of classification is obtained
using k-NN classifier for various k values for a fixed feature dimensionality of D=4. Figure 5.6 shows the
plot of average rate of classification and average number of classes in the set 𝐶>75 vs. various values of k.
0
0.2
0.4
0.6
0.8
1
1 2 3 4 5 6 7 8 9 10
Acc
ura
cy (
%)
Number of Nearest Neighbor
User A
0
0.2
0.4
0.6
0.8
1
1 2 3 4 5 6 7 8 9 10
Number of Nearest Neighbor
User B
0
0.2
0.4
0.6
0.8
1
1 2 3 4 5 6 7 8 9 10
Number of Nearest Neighbor
User C
0
0.2
0.4
0.6
0.8
1
1 2 3 4 5 6 7 8 9 10
Acc
ura
cy (
%)
Number of Nearest Neighbor
User D
0
0.2
0.4
0.6
0.8
1
1 2 3 4 5 6 7 8 9 10
Number of Nearest Neighbor
User E
0
0.2
0.4
0.6
0.8
1
1 2 3 4 5 6 7 8 9 10
Number of Nearest Neighbor
User F
63
Figure 5.6 True Rate of Classification for the entire Dataset for 1<k<10 (k ∈ 𝐼).
From the definition of TRC and Figure 5.6, we clearly see that for k=3, both the average rate of classification
and the number of classes in the set 𝐶>75 are higher as compared to other k values. This implies that for this
dataset, 3-Nearest Neighbor classifier is capable of correctly classifying maximum number of classes of
thumb motions with a higher average accuracy.
Table 5-3 shows the classification accuracies obtained using 3-Nearest Neighbor algorithm applied to each
user on a feature dimensionality of D = 4. Similar to QDA, pattern classification accuracy was highly user
dependent as shown in the user-wise distribution in Figure 5.7, being highest for User E and consistently
low for User C. However, as seen from this table, User C movements could be classified at better accuracy
compared to what was reported in Table 5-1 which shows the results using QDA algorithm.
0
1
2
3
4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 2 3 4 5 6 7 8 9 10
Nu
mb
er o
f C
lass
es C
orr
ectl
y C
lass
ifie
d
Cla
ssif
icat
ion
Acc
ura
cy (
%)
k Values for k-Nearest Neighbor
True Classification Rates for kNN
Avg. Classification Number of Correct Classes
64
User
PCA Reduced Feature Set (D=4)
No. of Movements
Avg. Classification Accuracy in % (Mean ± SD)
No. of Movements Discernible at ≥75%
Adjusted Accuracy in % (Mean ± SD)
A 4 69.5 ± 31.9 3 85.2 ± 6.4
B 4 90.0 ± 11.5 4 90.0 ± 11.5
C 4 71.87 ± 15.7 3 79.16 ± 7.2
D 4 79.5 ± 13.6 2 90.9 ± 0
E 4 93.7 ± 7.9 4 93.8 ± 7.9
F 4 81.25 ± 16.1 3 87.5 ± 12.5
Avg. (A-F) 4 80.9 ± 0.087 3.16 ± 0.75
(3.2 ± 0.83)b
87.75 ± 0.051
(89.4 ± 0.049)b
Table 5.3 Participant-wise Classification Accuracies using PCA and k-NN Algorithm
b – Classification accuracy excluding User C for comparing with QDA Algorithm.
Figure 5.7 shows the classification accuracies for each user for each thumb movement grouped per
movement class using k-NN algorithm. Data from Table 5-3 has been used for generating these plots, where
the classification accuracies obtained using 3-NN algorithm for all users are noted.
Contrary to the expectations set due to choices made at the experiment design phase and their validation
from the results obtained using QDA algorithm that thumb abduction (AB) is the least classifiable
movement, from Figure 5.7, we see that the spread of classification accuracies for the motion group thumb-
adduction (AD) is more scattered for most of the users. This implies that this motion is the most uncertain
for classification using kNN classifier.
65
Figure 5.7 Classification Accuracy per-User per-Class employing k-NN Algorithm
However, we can still support this result by arguing that none of two the mmgSensors were placed directly
on the muscles responsible for thumb adduction (AD), and hence we get these results. For User-C, User-D
and User-F, this motion could not be classified with a minimum classification accuracy of 75% and hence
this motion did not appear in the set 𝐶>75 for final classification accuracy calculations using the TRC metric.
5.3 Discussion
5.3.1 Discussion on Final Results
From the results section above, we see that PCA is an appropriate feature reduction algorithm and can be
used for feature reduction of MMG dataset for pattern recognition problems such as this. However, it is not
guaranteed to produce good classification accuracy results for all users, as seen in the case of User C. Fisher
score based feature pre-selection helps in some cases such as the case with User C, where its performance
improved considerably, but for others it did not have any effect on classification accuracy. It should be
noted that classification results using Fisher Score based pre-screening on features has not been shown in
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A D D U C T I O N A B D U C T I O N F L E X I O N E X T E N S I O N
Cla
ssiF
icat
ion
acc
ura
cy (
%)
Thumb Movement Types
C L A S S W I S E C L A S S I F I C A T I O N A C C U R A C Y P ER U S ER
User-A User-B User-C User-D User-E User-F
66
the final compilation as we could not find a fixed common set of algorithmic parameters for feature
selection that will give best results across all users. Hence for better feature selection, some advanced multi-
objective optimization algorithms such as genetic algorithms can be used, but such an implementation is a
scope for future work. Table 5-4 presents the summary of pattern classification accuracies using both QDA
and k-Nearest Neighbor algorithms on both the metrics namely Classification Accuracy and True Rate of
Classification (TRC).
Feature Selection Algorithm
Classification Algorithm
No. of Movements
Avg. Classification Accuracy in % (Mean ± SD)
No. of Movements Discernible
at ≥75%
Adjusted Accuracy in % (Mean ± SD)
PCA QDA 4 81.5 ± 0.12 3.2 ± 0.84 90.5 ± 0.041
PCA kNN 4 80.9 ± 0.087 3.2 ± 0.83 89.4 ± 0.049
Table 5.4 Summary of Results for 2 Classifiers using the Two Metrics Described in Section 5.2.1
As can be inferred from Table 5-4, QDA out performs k-NN algorithm on both the metrics described in
Section 5.2.1. Using the TRC metric, QDA can clearly predict more number of classes of thumb motion at
a better classification rate, more often as compared to k-NN classifier. Thus, QDA can be reliably used for
MMG pattern recognition problems. However, it can be seen here that the difference of average accuracies
between the two classifiers on both the metrics is not high. It should also be noted that k-NN classifier
performs considerably better for User C and predicts at a classification accuracy of 71% as compared to
59% using QDA on the same PCA based reduced feature set of D=4. This may be used to infer that k-NN
classifier is more robust in tackling user data and can predict with more accuracy on diverse types of user
cases.
QDA is computationally more intensive compared to k-NN and may not suit the real-time applications
scenario such as the design of computer interfaces, in which case, k-NN may prove to be useful. Literature
review shows that there are various data-structures such as kd-trees or ball-trees [60, 61] that can be used
for computationally faster implementation of k-NN, but such an implementation is a scope for future work.
67
5.3.2 Discussion on User C
One of the short-comings of the wristband design was that the material used to make the band did not have
resizability. Having different sizes for different user wrists was attained using various positions on the
Velcro tie material used to tie the bands. One distinguishing observation for User C was sensor fitting issues
i.e., exact positioning of the sensors above the muscles could not be attained, due to the bigger
circumference of the wrist. Inspection of the time series data for User C shows lower energy signals for
each of the movements compared to other users. Similarly, another observation relevant to User C, User D
and User F are that their movements were unstable and the thumb was shaky during the sample captures.
This made data acquisition difficult for these users. Although, User D and User F have better classification
accuracies compared to User C, they have lower accuracies compared to other users.
The pattern classification results on User C obtained by employing QDA on PCA based dimensionality
reduction as shown in Table 5-1 show that User C is clearly an anomaly among other users with similar
dimensionality of D=4. This is further verified by the best case classification accuracy results for User C
noted in Table 5-2, that User C gives best classification results at D=2, in which case all 4 motions appear
in the set 𝐶>75. Thus for the final comparison between the algorithms as discussed in the next section, we
will report results from the entire dataset when compiling results using average classification accuracy
metric. However, we did not include User-C results while comparing the algorithms using the TRC metric,
as shown in Table 5-4.
68
Chapter 6 - Conclusion and Future Work
6.1 Conclusion
A mechanomyography based wearable thumb joystick has been integrated into a wristband to use as a
replacement of Nunchuck joystick in live performances by Virginia Tech’s Linux Laptop Orchestra (L2-
Ork). Six voluntary participants took part in a user study to characterize the performance of the designed
user interface. During the user study, continuous time-series mechanomyographic data were captured from
each user after carefully placing the wristband at desired location on the forearm.
The results of two pattern recognition algorithms were compared for classification accuracy on four classes
of thumb motion. The same algorithms were also compared on a new metric defined during the course of
analysis, which we have named as true-rate of classification (TRC). This metric gives a measure of the
number of movement classes that can be reliably classified above a minimum threshold of 75% accuracy.
Using both of these metrics, quadratic discriminant analysis (QDA) classifier outperforms k-nearest
neighbor classifier (k-NN), although by a small margin.
The wearable interface designed here could correctly classify 81.5% of all four thumb motions on an
average across all users employing QDA classifier. Using the TRC metric, we can also say that our interface
can correctly classify three motion classes with an accuracy of 90.5% across all users employing QDA
classifier. However, k-NN classifier is more robust to varied types of user data such as User-C whose
gesture data is clearly an outlier due to known sensor fitting issues as discussed in Section 5.3.2. In spite of
that, k-NN classifier gives better classification accuracy results in case of the outlier User-C as compared
to QDA classifier.
The TRC metric was used to identify the most commonly misclassified thumb motions. It was found that
thumb abduction (AB) and thumb adduction (AD) were the most commonly misclassified motion classes.
We postulate that this may be due to the fact that the muscles responsible for these motions did not have
69
mmgSensors placed directly above them which might have resulted in poorer data for these classes of
motion.
While these results do not ensure a completely reliable design of a user interface, but an overall
classification accuracy of 81.5% and a TRC based accuracy of 90.5% for 3 classes of thumb motion
achieved using only two mmgSensors indicate that mechanomyography based wearable thumb-joystick is
a feasible design idea worthy of further study.
6.2 Future Work
The most compelling area of further work is to verify whether placing mmgSensors directly above the
muscles responsible for thumb abduction (AB) and thumb adduction (AD) really improves the overall
classification accuracy, thereby making it a reliable wearable interface. At present, this could not be done
due to the limitations of the data acquisition setup which allows only two channels of sound data to be
captured. In addition, an elastic band should be used to design the wristband thereby making it to be user
wrist-size agnostic.
The next logically progressive step will be to package the system as a real-time gesture recognizer. Wired
interfaces tend to have limited usability among users, more so when it is targeted for an artistic use such as
in our use case scenario of L2-Ork. For such cases, developing wireless data transmission from the
wristband using Blue-tooth Low Energy modules can also be looked into.
Use of mechanomyographic signal as a part of a multimodal gesture recognition system where two or more
types of sensors are used to create a sensor data fusion, is an exciting way forward from here. A startup
called ThalmicLabs from Toronto Canada is working on a similar idea using undisclosed muscle machine
sensors along with inertial sensors to create a gesture recognition armband named [62]. Our work is in the
same vein and aims at a very similar application. Recent work in this area proves this should be an
interesting way forward.
70
Bibliography
[1] S. F. George Tzanetakis. (2013, 11/25/2013). Tutorial 15: Multi-modal user interfaces: a new music instruments perspective. Available: http://www.icassp2013.com/Tutorial_15.asp
[2] D. Rosenboom, Biofeedback and the Arts: Results of Early Experiments: Aesthetic Research Centre of Canada, 1974.
[3] M. Donnarumma, "Music for Flesh II: informing interactive music performance with the viscerality of the body system," in Proceedings of the 2012 Conference on New Interfaces for Musical Expression (NIME-12), 2012.
[4] R. B. Knapp and H. S. Lusted, "A bioelectric controller for computer music applications," Computer music journal, vol. 14, pp. 42-47, 1990.
[5] G. Wang. (2008, 11/25/2013). Stanford Laptop Orchestra. Available: http://slork.stanford.edu/bio/
[6] I. Bukvic. (2009, 11/25/2013). Linux Laptop Orchestra. Available: http://l2ork.music.vt.edu/main/ [7] L2-Ork. (2013, 11/25/2013). What differentiates L2Ork from other *Orks? Available:
http://l2ork.music.vt.edu/main/?page_id=63 [8] J. B. Schaefer. (2013, 11/25/2013). Virginia Tech laptop orchestra joins Boyer for unique
performance. Available: http://news.temple.edu/news/2013-04-18/l2ork-laptop-orchestra-perform-members-temple%E2%80%99s-boyer-college-music-and-dance
[9] T. Crisp. (2010, 11/25/2013). Approaches to Liberation Liberating the Body : Chapter Four. Available: http://dreamhawk.com/body-and-mind/approaches-to-liberation/
[10] J. Bae and R. M. Voyles, "Wearable joystick for gloves-on human/computer interaction," in Defense and Security Symposium, 2006, pp. 62300N-62300N-10.
[11] V. R. Ltd. (2013, 11/25/2013). Data Gloves. Available: http://www.vrealities.com/products/data-gloves
[12] A. Posatskiy, "Design and Evaluation of Pressure-based Sensors for Mechanomyography: an Investigation of Chamber Geometry and Motion Artifact," 2011.
[13] M. A. Islam, K. Sundaraj, R. B. Ahmad, N. U. Ahamed, and M. Babu, "Mechanomyography Sensor Development, Related Signal Processing and Applications: A Systematic Review," 2013.
[14] M. Ma, "MMG sensor for muscle activity detection: Low cost design, implementation and experimentation," 2010.
[15] K. Kalgaonkar and B. Raj, "One-handed gesture recognition using ultrasonic Doppler sonar," in Acoustics, Speech and Signal Processing, 2009. ICASSP 2009. IEEE International Conference on, 2009, pp. 1889-1892.
[16] D. Kim, O. Hilliges, S. Izadi, A. D. Butler, J. Chen, I. Oikonomidis, et al., "Digits: freehand 3D interactions anywhere using a wrist-worn gloveless sensor," in Proceedings of the 25th annual ACM symposium on User interface software and technology, 2012, pp. 167-176.
[17] S. Gustafson, D. Bierwirth, and P. Baudisch, "Imaginary interfaces: spatial interaction with empty hands and without visual feedback," in Proceedings of the 23nd annual ACM symposium on User interface software and technology, 2010, pp. 3-12.
[18] T. Lee and T. Hollerer, "Handy AR: Markerless inspection of augmented reality objects using fingertip tracking," in Wearable Computers, 2007 11th IEEE International Symposium on, 2007, pp. 83-90.
[19] B. Buxton, "Multi-touch systems that i have known and loved," Microsoft Research, pp. 1-10, 2007.
71
[20] I. Oikonomidis, N. Kyriazis, and A. A. Argyros, "Full DOF tracking of a hand interacting with an object by modeling occlusions and physical constraints," in Computer Vision (ICCV), 2011 IEEE International Conference on, 2011, pp. 2088-2095.
[21] J. Romero, H. Kjellstrom, and D. Kragic, "Hands in action: Real-time 3D reconstruction of hands in interaction with objects," in Robotics and Automation (ICRA), 2010 IEEE International Conference on, 2010, pp. 458-463.
[22] K. D. Nguyen, I.-M. Chen, Z. Luo, S. H. Yeo, and H.-L. Duh, "A wearable sensing system for tracking and monitoring of functional arm movement," Mechatronics, IEEE/ASME Transactions on, vol. 16, pp. 213-220, 2011.
[23] D. J. Ryan. Finger and gesture recognition with Microsoft Kinect. [24] S. Gupta, D. Morris, S. Patel, and D. Tan, "Soundwave: using the doppler effect to sense gestures,"
in Proceedings of the 2012 ACM annual conference on Human Factors in Computing Systems, 2012, pp. 1911-1914.
[25] C. Microsystems. (2013, 11/25/2013). Chirp Ultrasound Gesture Recogniser. Available: http://chirpmicro.com/
[26] T. S. Saponas, D. S. Tan, D. Morris, R. Balakrishnan, J. Turner, and J. A. Landay, "Enabling always-available input with muscle-computer interfaces," in Proceedings of the 22nd annual ACM symposium on User interface software and technology, 2009, pp. 167-176.
[27] N. Alves and T. Chau, "Uncovering patterns of forearm muscle activity using multi-channel mechanomyography," Journal of electromyography and kinesiology, vol. 20, pp. 777-786, 2010.
[28] N. L. Alves-Kotzev, "Mechanomyography as an Access Pathway for Binary and Multifunction Control," University of Toronto, 2010.
[29] C. Orizio, "Muscle sound: bases for the introduction of a mechanomyographic signal in muscle studies," Critical reviews in biomedical engineering, vol. 21, pp. 201-243, 1992.
[30] J. Mercer, N. Bezodis, D. DeLion, T. Zachry, and M. Rubley, "EMG sensor location: Does it influence the ability to detect differences in muscle contraction conditions?," Journal of Electromyography and Kinesiology, vol. 16, pp. 198-204, 2006.
[31] A. J. Fuglevand, D. A. Winter, A. E. Patla, and D. Stashuk, "Detection of motor unit action potentials with surface electrodes: influence of electrode size and spacing," Biological Cybernetics, vol. 67, pp. 143-153, 1992.
[32] R.-F. D. Galiana-Merino JJ, Martinez-Espla JJ., "Power line interference filtering on surface electromyography based on the stationary wavelet packet transform," Elsevier Journal in Computer Methods and Programs in Biomedicine, 2013.
[33] A. Posatskiy and T. Chau, "Design and evaluation of a novel microphone-based mechanomyography sensor with cylindrical and conical acoustic chambers," Medical engineering & physics, vol. 34, pp. 1184-1190, 2012.
[34] M. Watakabe, K. Mita, K. Akataki, and Y. Itoh, "Mechanical behaviour of condenser microphone in mechanomyography," Medical and biological engineering and computing, vol. 39, pp. 195-201, 2001.
[35] T. W. Beck, Applications of Mechanomyography for examining muscle function, 2010. [36] C. Orizio, D. Liberati, C. Locatelli, D. De Grandis, and A. Veicsteinas, "Surface mechanomyogram
reflects muscle fibres twitches summation," Journal of biomechanics, vol. 29, pp. 475-481, 1996. [37] D. T. Barry and N. M. Cole, "Muscle sounds are emitted at the resonant frequencies of skeletal
muscle," Biomedical Engineering, IEEE Transactions on, vol. 37, pp. 525-531, 1990. [38] J. V. Frangioni, T. S. Kwan-Gett, L. E. Dobrunz, and T. A. McMahon, "The mechanism of low-
frequency sound production in muscle," Biophysical journal, vol. 51, pp. 775-783, 1987.
72
[39] A. Courteville, T. Gharbi, and J.-Y. Cornu, "MMG measurement: a high-sensitivity microphone-based sensor for clinical use," Biomedical Engineering, IEEE Transactions on, vol. 45, pp. 145-150, 1998.
[40] G. Oster and J. S. Jaffe, "Low frequency sounds from sustained contraction of human skeletal muscle," Biophysical journal, vol. 30, pp. 119-127, 1980.
[41] D. T. Barry, S. R. Geiringer, and R. D. Ball, "Acoustic myography: a noninvasive monitor of motor unit fatigue," Muscle & nerve, vol. 8, pp. 189-194, 1985.
[42] J. Silva, W. Heim, and T. Chau, "A self-contained, mechanomyography-driven externally powered prosthesis," Archives of physical medicine and rehabilitation, vol. 86, pp. 2066-2070, 2005.
[43] J. Silva, W. Heim, and T. Chau, "MMG-based classification of muscle activity for prosthesis control," in Engineering in Medicine and Biology Society, 2004. IEMBS'04. 26th Annual International Conference of the IEEE, 2004, pp. 968-971.
[44] N. Alves and T. Chau, "The design and testing of a novel mechanomyogram-driven switch controlled by small eyebrow movements," Journal of NeuroEngineering and Rehabilitation, vol. 7, p. 22, 2010.
[45] J. Silva and T. Chau, "Coupled microphone-accelerometer sensor pair for dynamic noise reduction in MMG signal recording," Electronics Letters, vol. 39, pp. 1496-1498, 2003.
[46] Y. Zeng, Z. Yang, W. Cao, and C. Xia, "Hand-motion patterns recognition based on mechanomyographic signal analysis," in BioMedical Information Engineering, 2009. FBIE 2009. International Conference on Future, 2009, pp. 21-24.
[47] A. Grossman, J. Silva, and T. Chauz, "Functional mapping of multiple mechanomyographic signals to hand kinematics," in Electrical and Computer Engineering, 2004. Canadian Conference on, 2004, pp. 493-496.
[48] W. P. Cooney, M. J. Lucca, E. Chao, and R. Linscheid, "The kinesiology of the thumb trapeziometacarpal joint," J Bone Joint Surg Am, vol. 63, pp. 1371-81, 1981.
[49] A. Hollister, W. L. Buford, L. M. Myers, D. J. Giurintano, and A. Novick, "The axes of rotation of the thumb carpometacarpal joint," Journal of Orthopaedic Research, vol. 10, pp. 454-460, 1992.
[50] J. F. Kramer, "Determination of thumb position using measurements of abduction and rotation," ed: Google Patents, 1996.
[51] F. M. Ronan O'Rahilly, Stanley Carpenter,Rand Swenson,. (2008, 11/25/2013). Basic Human Anatomy, Chapter 10. Available: http://www.dartmouth.edu/~humananatomy/part_2/chapter_11.html
[52] D. Hosford. (1998, 11/25/2013). The Hosford Muscle Tables: Skeletal Muscles of the Human Body. Available: http://www.ptcentral.com/muscles/musclearms.html
[53] E. n. LLC. (2013, 11/25/2013). Thumb Articulations. Available: http://www.exrx.net/Articulations/Thumb.html
[54] A. Phinyomark, C. Limsakul, and P. Phukpattaranont, "A novel feature extraction for robust EMG pattern recognition," arXiv preprint arXiv:0912.3973, 2009.
[55] J. Lee, C. M. Steele, and T. Chau, "Classification of healthy and abnormal swallows based on accelerometry and nasal airflow signals," Artificial Intelligence in Medicine, vol. 52, pp. 17-25, 2011.
[56] D. L. Donoho, "High-dimensional data analysis: The curses and blessings of dimensionality," AMS Math Challenges Lecture, pp. 1-32, 2000.
[57] N. Kouiroukidis and G. Evangelidis, "The Effects of Dimensionality Curse in High Dimensional kNN Search," in Informatics (PCI), 2011 15th Panhellenic Conference on, 2011, pp. 41-45.
[58] D. H. Jeong, C. Ziemkiewicz, W. Ribarsky, R. Chang, and C. V. Center, "Understanding Principal Component Analysis Using a Visual Analytics Tool."
73
[59] T. P. S. University. (2013, 11/25/2013). Applied Data Mining. Available: https://onlinecourses.science.psu.edu/stat857/
[60] P. Piro, F. Nielsen, and M. Barlaud, "Tailored bregman ball trees for effective nearest neighbors," in European Workshop on Computational Geometry (EuroCG), LORIA, Nancy, France, 2009.
[61] N. Bhatia, "Survey of nearest neighbor techniques," arXiv preprint arXiv:1007.0085, 2010. [62] T. Labs. (2013, 11/25/2013). Myo Gesture Recognition Wristband. Available:
https://www.thalmic.com/en/myo/